Notice of Pre-AIA or AIA Status
The present application, filed on or after March 16, 2013, is being examined under the first inventor to file provisions of the AIA .
DETAILED ACTION
This action is responsive to the Application filed on 05/18/2026
Claims 1-19 and 26 are pending in the case. Claims 1, 14 and 26 are independent claims. Claims 1-2, 14-15 and 26 have been previously amended. Claims 20-25 and 27 have been canceled.
Claim Rejections - 35 USC § 103
The following is a quotation of 35 U.S.C. 103 which forms the basis for all obviousness rejections set forth in this Office action:
A patent for a claimed invention may not be obtained, notwithstanding that the claimed invention is not identically disclosed as set forth in section 102, if the differences between the claimed invention and the prior art are such that the claimed invention as a whole would have been obvious before the effective filing date of the claimed invention to a person having ordinary skill in the art to which the claimed invention pertains. Patentability shall not be negated by the manner in which the invention was made.
Claims 1-9 and 11-18 are rejected under 35 U.S.C. 103 as being unpatentable over “Robust Road Attribute Inference With Graph Neural Networks”, https://arxiv.org/pdf/1912.12408, Songtao et al, 12/28/2019, hereinafter referred to as Songtao, in view of “Predicting Transportation Modes of GPS Trajectories using Feature Engineering and Noise Removal”, https://arxiv.org/pdf/1802.10164, Etemad et al, 02/27/2028, hereinafter referred to as Etemad.
With respect to claim 1, Songtao disclose:
A method of predicting one or more road attributes corresponding to roads in a geographical area, the geographical area comprising road segments, the method comprising (On page 1 and Fig. 1, Songtao disclose predicted road attributes. In Fig. 1, Songtao discloses a geographic area represented as a road network graph comprising connected road portions between graph vertices. These connected portions correspond to road segments within the geographical area.)
Providing map data, wherein the map data comprises image data of the geographical area (On page 4 and Fig. 3, Songtao discloses providing map data comprising image data of a geographical area. Specifically, the reference receives satellite imagery depicting the geographical area together with a corresponding road network graph as input to the neural-network framework. The satellite imagery consists of image data of the geographical area, while the road network graph provides additional map information.)
Extracting map features from the map data (On page 3(3 RoadTagger), Songtao discloses extracting map features from map data. Specifically, it receives satellite imagery of a geographical area as map data and applies a CNN encoder to derive a feature vector from the satellite imagery at each road-network vertex. The derived feature vector represents features extracted from the map image and is subsequently provided to the graph neural network for road-attribute inference.)
Using at least one processor for: predicting road attributes by inputting the (On page 3 (3 RoadTagger), Songtao discloses inputting features into a neural network. Specifically, the CNN encoder into satellite imagery to extract a 64-dimensional embedding (feature vector) for each road-network vertex. The extracted embeddings are then provided as inputs to a Graph Neural Network (GNN), which propagates the feature information over a road network graph to generate road-attribute prediction. On page 5 (4 Evaluation), Songtao discloses the image classifier, the CNN extracts image/map features and the GNN performs the final inference/classification of road attributes. (On page 6, paragraph 11, Songtao disclose different visual features (e.g., bridges, trees, intersections, etc.) on road attributes.))
A second sub-neural network for the map features (On page 5 (4 Evaluation), Songtao discloses the image classifier, the CNN extracts image/map features and the GNN performs the final inference/classification of road attributes. On page 7, Songtao disclosed raw for the original road network graph, road for the road extraction graph, Road(D) for the road extraction graph with directional decomposition and Aux for the auxiliary graph for parallel roads.)
With respect to claim 1, Songtao do not explicitly disclose:
Extracting trajectory features from the trajectory data
Trajectory features
Providing trajectory data of the geographical area, the trajectory data received as raw trace data recorded by one or more vehicles traversing at least a portion of the geographical area and including at least one of location, bearing and speed data
Classifying the output of the neural network into prediction probabilities of the road attributes, wherein the neural network comprises a first sub-neural network for the trajectory features and
However, it is known by Etemad to disclose:
Extracting trajectory features from the trajectory data (On page 3, Etemad disclose calculating all the point features for each trajectory, extracting some
statistical attributes referred to as trajectory features.)
Trajectory features (On page 3, Etemad discloses trajectory features are divided into two different types: (i) global trajectory features, which summarize information regarding the whole trajectory in a single value; and (ii) local trajectory features, which describe a local part of the trajectory.)
Providing trajectory data of the geographical area, the trajectory data received as raw trace data recorded by one or more vehicles traversing at least a portion of the geographical area and including at least one of location, bearing and speed data (On page 2 (A framework for transportation mode prediction), Etemad discloses calculating estimated speed and bearing, and later identifies point features including distance rate, speed, acceleration, bearing, jerking, bearing rate, and rate of bearing rate. In Fig. 1 and page 3, Etemad discloses raw GPS data points (e.g., latitude, longitude and time, calls a trajectory a sequence of GPS points, and Figure 1 expressly begins with Raw Trajectories. It further explains that the GPS point represents a moving object's movement and that trajectories are generated from the raw GPS point.))
Classifying the output of the neural network into prediction probabilities of the road attributes, wherein the neural network comprises a first sub-neural network for the trajectory features (On page 2 and Fig. 1, Etemad discloses classifying research in transportation modes prediction regarding the type of features in two branches: (i) domain expert features; and (ii) learned features. From raw GPS data points (e.g., latitude, longitude and time) it is possible to calculate many attributes regarding the moving object's movement.)
Songtao and Etemad are analogous pieces of art because both solve related transportation-machine-learning problems using feature extraction and classification. Accordingly, it would have been obvious to a person of ordinary skill in the art, before the effective filing date of the claimed invention, to modify Songtao, with extracting semantic information from map imagery as taught by Songtao, with the GPS trajectory paper teaches extracting trajectory features from movement data as taught by Etemad. The motivation for doing so would have been to improve inference accuracy over the CNN image classifier-based approaches (See (On page 1 (Abstract) of Songtao.)
Regarding claim 2, Songtao in view of Etemad and Pack disclose the elements of claim 1. In addition, Songtao disclose:
The method of claim 1, wherein the neural network is configured to receive the trajectory features and the map features, and wherein classifying is executed by a classifier, the classifier being configured to calculate one or more of the prediction probabilities based on the task-specific fused representations (On page 4 and Fig. 3, Songtao discloses the fully connected layers act as the claimed classifier. The learned embedding (task-specific fused representation) produced by the CNN and GNN is input to the classifier, which computes prediction probabilities using Softmax before outputting the final road attribute prediction.)
Regarding claim 3, Songtao in view of Etemad disclose the elements of claim 1. In addition, Songtao disclose:
The method of claim 1, wherein the trajectory features are processed by the first sub-neural network into shared global trajectory features, wherein the first sub- neural network comprises one or more fully-connected layers (On page 3 (RoadTagger), Songtao discloses that the CNN encoder uses 12 convolutional layers and 3 fully-connected layers to extract a 64-dimension embedding for each vertex.)
Regarding claim 6, Songtao, in view of Etemad discloses the elements of claim 1. In addition, Songtao disclose:
The method of claim 1, wherein the map features are processed by the second sub-neural network into shared global map features (On page 3 (RoadTagger), Songtao discloses the map processing features by a second sub-neural network into shared global map features. Specifically, a CNN derives feature vectors from satellite imagery corresponding to road-network vertices, and the resulting map-image features are supplied to a GNN. The GUN propagates and aggregates the features along the road-network graph, such that the representation at a vertex incorporates feature information from other portions of the road.)
Regarding claim 10, Songtao, in view of Etemad disclose the elements of claim 1. In addition, Songtao disclose:
The method of claim 1, wherein extracting map features from the map data comprises generating cropped images by cropping images from the image data, wherein the cropped images are centered at a corresponding road segment of the road segments (On page 3 (RoadTagger), Songtao discloses inputting satellite imagery together with a road network graph, cropping satellite images around each location in the road graph. Each group is centered on the corresponding road location so that the CNN can extract image features. )
Regarding claim 11, Songtao, in view of Etemad discloses the elements of claim 1. In addition, Etemad disclose:
The method of claim 1, wherein extracting trajectory features from the trajectory data comprises determining group of traces of the trajectory data that are associated with a road segment of the road segments (On page 3 (first paragraph), Etemad discloses extracting statistical attributes referred to as trajectory features.)
Regarding claim 12, Songtao, in view of Etemad disclose the elements of claim 11. In addition, Etemad disclose:
The method of claim 11, wherein extracting trajectory features from the trajectory data further comprises calculating respective distributions of one or more of location, bearing, and speed, and using the distributions as the trajectory features (On page 3 (first paragraph), Etemad discloses extracting statistical attributes referred to as trajectory features.)
Regarding claim 13, Songtao in view of Etemad disclose the elements of claim 1. In addition, Etemad disclose:
The method of claim 1, wherein the trajectory data comprises a plurality of data points, each data point comprising latitude, longitude, bearing, and speed (On page 2, Etemad discloses the system computes movement-related attributes such as estimated speed and bearing (direction of travel.))
With respect to claim 14, Songtao disclose:
A data processing system comprising one or more processors configured to carry out predicting road attributes corresponding to roads in a geographical area, comprising (On page 1 and Fig. 1, Songtao disclose predicted road attributes. In Fig. 1, Songtao discloses a geographic area represented as a road network graph comprising connected road portions between graph vertices. These connected portions correspond to road segments within the geographical area.)
A second memory configured to store map data, wherein the map data comprises image data of the geographical area (On page 4 and Fig. 3, Songtao discloses providing map data comprising image data of a geographical area. Specifically, the reference receives satellite imagery depicting the geographical area together with a corresponding road network graph as input to the neural-network framework. The satellite imagery consists of image data of the geographical area, while the road network graph provides additional map information.)
A map feature extractor configured to extract map features from the map data (On page 3(3 RoadTagger), Songtao discloses extracting map features from map data. Specifically, it receives satellite imagery of a geographical area as map data and applies a CNN encoder to derive a feature vector from the satellite imagery at each road-network vertex. The derived feature vector represents features extracted from the map image and is subsequently provided to the graph neural network for road-attribute inference. (On page 5 (4 Evaluation), Songtao discloses the image classifier, the CNN extracts image/map features and the GNN performs the final inference/classification of road attributes.))
A neural network configured to predict road attributes by generating an output of task-specific fused representations based on the (On page 3 (3 RoadTagger), Songtao discloses inputting features into a neural network. Specifically, the CNN encoder into satellite imagery to extract a 64-dimensional embedding (feature vector) for each road-network vertex. The extracted embeddings are then provided as inputs to a Graph Neural Network (GNN), which propagates the feature information over a road network graph to generate road-attribute prediction. On page 5 (4 Evaluation), Songtao discloses the image classifier, the CNN extracts image/map features and the GNN performs the final inference/classification of road attributes. (On page 6, paragraph 11, Songtao disclose different visual features (e.g., bridges, trees, intersections, etc.) on road attributes.))
A second sub-neural network for the map features (On page 5 (4 Evaluation), Songtao discloses the image classifier, the CNN extracts image/map features and the GNN performs the final inference/classification of road attributes. On page 7, Songtao disclosed raw for the original road network graph, road for the road extraction graph, Road(D) for the road extraction graph with directional decomposition and Aux for the auxiliary graph for parallel roads.)
With respect to claim 14, Songtao do not explicitly disclose:
A first memory configured to store trajectory data of the geographical area, the trajectory data received as raw trace data recorded by one or more vehicles traversing at least a portion of the geographical area and including at least one of location, bearing and speed data
A trajectory feature extractor configured to extract trajectory features from the trajectory data
Trajectory features
A classifier configured to classify the output of the neural network into prediction probabilities of the road attributes, wherein the neural network comprises a first sub-neural network for the trajectory features
However, it is known by Etemad to disclose:
A first memory configured to store trajectory data of the geographical area, the trajectory data received as raw trace data recorded by one or more vehicles traversing at least a portion of the geographical area and including at least one of location, bearing and speed data (On page 2 (A framework for transportation mode prediction), Etemad discloses calculating estimated speed and bearing, and later identifies point features including distance rate, speed, acceleration, bearing, jerking, bearing rate, and rate of bearing rate. In Fig. 1 and page 3, Etemad discloses raw GPS data points (e.g., latitude, longitude and time, calls a trajectory a sequence of GPS points, and Figure 1 expressly begins with Raw Trajectories. It further explains that the GPS point represents a moving object's movement and that trajectories are generated from the raw GPS point.))
A trajectory feature extractor configured to extract trajectory features from the trajectory data (On page 3, Etemad discloses calculating all the point features for each trajectory, extracting some statistical attributes referred to as trajectory features.)
Trajectory features (On page 3, Etemad discloses trajectory features are divided into two different types: (i) global trajectory features, which summarize information regarding the whole trajectory in a single value; and (ii) local trajectory features, which describe a local part of the trajectory.)
Classifying the output of the neural network into prediction probabilities of the road attributes, wherein the neural network comprises a first sub-neural network for the trajectory features (On page 2 and Fig. 1, Etemad discloses classifying research in transportation modes prediction regarding the type of features in two branches: (i) domain expert features; and (ii) learned features. From raw GPS data points (e.g., latitude, longitude and time) it is possible to calculate many attributes regarding the moving object's movement.)
Songtao and Etemad are analogous pieces of art because both solve related transportation-machine-learning problems using feature extraction and classification. Accordingly, it would have been obvious to a person of ordinary skill in the art, before the effective filing date of the claimed invention, to modify Songtao, with extracting semantic information from map imagery as taught by Songtao, with the GPS trajectory paper teaches extracting trajectory features from movement data as taught by Etemad. The motivation for doing so would have been to improve inference accuracy over the CNN image classifier-based approaches (See (On page 1 (Abstract) of Songtao.)
Regarding claim 15, Songtao in view of Etemad disclose the elements of claim 14. In addition, Songtao disclose:
The data processing system of claim 14, wherein the neural network is configured to receive the trajectory features and the map features, and wherein the classifier is configured to calculate one or more of the prediction probabilities based on the task-specific fused representations (On page 4 and Fig. 3, Songtao discloses the fully connected layers act as the claimed classifier. The learned embedding (task-specific fused representation) produced by the CNN and GNN is input to the classifier, which computes prediction probabilities using Softmax before outputting the final road attribute prediction.)
Regarding claim 16, Songtao, in view of Etemad disclose the elements of claim 15. In addition, Songtao disclose:
The data processing system of claim 15, wherein the first sub- neural network is configured to process the trajectory features into shared global trajectory features, wherein the first sub-neural network comprises one or more fully-connected layers (On page 3 (RoadTagger), Songtao discloses that the CNN encoder uses 12 convolutional layers and 3 fully-connected layers to extract a 64-dimension embedding for each vertex.)
Regarding claim 18, Songtao in view of Etemad disclose the elements of claim 17. In addition, Songtao disclose:
The data processing system of claim 17, wherein the first sub- neural network is configured to process the trajectory features into shared global trajectory features, wherein the first sub-neural network comprises one or more fully-connected layers (On page 3 (RoadTagger), Songtao discloses that the CNN encoder uses 12 convolutional layers and 3 fully-connected layers to extract a 64-dimension embedding for each vertex.)
Claims 4-5, 7-9, 17 and 19 are rejected under 35 U.S.C. 103 as being unpatentable over Songtao, in view of Etemad and further in view of Shazeer et al. (US Patent No. 10,452,978 B2) hereinafter referred to as Shazeer.
Regarding claim 4, Songtao in view of Etemad disclose elements of claim 2. Songtao in view of Etemad do not disclose:
The method of claim 2, further comprising determining attention scores of pre-defined indicators corresponding to road attributes based on the trajectory data, wherein the pre-defined indicators are processed by a fully connected layer, and wherein the attention scores are determined based on activation functions
However, Shazeer disclose the limitation (In Col. 5, lines 35–57, Shazeer discloses using an activation function (e.g., Rule) within the neural network feed-forward layer. The encoder subnetwork may also include a connection layer.)
Accordingly, it would have been obvious to a person of ordinary skill in the art, before the effective filing date of the claimed invention, having the teaching of Songtao in view of Etemad to include Shazeer, with an attention mechanism obtain the weights on the values. The attention sub-layer then computes a weighted sum of the values in accordance with these weights as taught by Shazeer. The motivation for doing so would have been to improve performance over conventional machine translation neural networks without task-specific tuning through the use of the attention mechanism (See (Col. 2, lines 28-30) of Shazeer.)
Regarding claim 5, Songtao in view of Etemad disclose elements of claim 4. Songtao in view of Etemad do not disclose:
The method of claim 4, wherein trajectory task-specific weighted representations are calculated based on the fusion of the attention scores with the shared global trajectory features of the first sub-neural network
However, Shazeer disclose the limitation (In Col. 4, lines 15–29, Shazeer discloses that the weighted vectors are combined (weighted sum) to generate a new context representation that is forwarded to later layers.)
Accordingly, it would have been obvious to a person of ordinary skill in the art, before the effective filing date of the claimed invention, having the teaching of Songtao in view of Etemad to include Shazeer, with an attention mechanism obtain the weights on the values. The attention sub-layer then computes a weighted sum of the values in accordance with these weights as taught by Shazeer. The motivation for doing so would have been to improve performance over conventional machine translation neural networks without task-specific tuning through the use of the attention mechanism (See (Col. 2, lines 28-30) of Shazeer.)
Regarding claim 7, Songtao in view of Etemad disclose elements of claim 6. Songtao in view of Etemad do not disclose:
The method of claim 6, further comprising calculating second attention scores of pre-defined indicators based on the shared global map features, wherein the pre-defined indicators are processed by a second fully connected layer, and wherein the second attention scores are determined based on activation functions
However, Shazeer disclose the limitation (In Col. 8, lines 38–50, Shazeer discloses multiple attention sublayers compute attention weights repeatedly throughout the network. Later attention layer operates on contextualized feature representation generated by the previous layer. Learned linear (fully connected) transformations are applied to queries, keys, and values before attention computation. The patent discloses feed-forward subnetworks using activation function (e.g., ReLU) together with the attention architecture.)
Accordingly, it would have been obvious to a person of ordinary skill in the art, before the effective filing date of the claimed invention, having the teaching of Songtao in view of Etemad to include Shazeer, with an attention mechanism obtain the weights on the values. The attention sub-layer then computes a weighted sum of the values in accordance with these weights as taught by Shazeer. The motivation for doing so would have been to improve performance over conventional machine translation neural networks without task-specific tuning through the use of the attention mechanism (See (Col. 2, lines 28-30) of Shazeer.)
Regarding claim 8, Songtao in view of Etemad disclose elements of claim 7. Songtao in view of Etemad do not disclose:
The method of claim 7, wherein map task-specific weighted representations are calculated based on the fusion of the second attention scores with the shared global map features of the second sub-neural network
However, Shazeer disclose the limitation (In Col.8, lines 38–50, Shazeer discloses later that attention sublayers compute another set of attention weights during a subsequent attention operation. Encoder subnetworks generate contextual encoded representation shared throughout the network. The attention mechanism computes attention weights and applies those weights to the encoded representation, producing a weighted output representation.)
Accordingly, it would have been obvious to a person of ordinary skill in the art, before the effective filing date of the claimed invention, having the teaching of Songtao in view of Etemad to include Shazeer, with an attention mechanism obtain the weights on the values. The attention sub-layer then computes a weighted sum of the values in accordance with these weights as taught by Shazeer. The motivation for doing so would have been to improve performance over conventional machine translation neural networks without task-specific tuning through the use of the attention mechanism (See (Col. 2, lines 28-30) of Shazeer.)
Regarding claim 9, Songtao in view of Etemad disclose elements of claim 5. Songtao in view of Etemad do not disclose:
The method of claim 5, wherein the map features are processed by the second sub-neural network into shared global map features
The method further comprises calculating second attention scores of pre-defined indicators based on the shared global map features, wherein the pre-defined indicators are processed by a second fully connected layer
The second attention scores are determined based on activation functions, wherein map task- specific weighted representations are calculated based on the fusion of the second attention scores with the shared global map features of the second sub-neural network
Task-specific fused representations are determined based on the map task-specific weighted representations and the trajectory task-specific weighted representations
However, Shazeer disclose the limitations:
The method of claim 5, wherein the map features are processed by the second sub-neural network into shared global map features (In Col. 8, lines 38–50, Shazeer discloses multiple attention sublayers compute attention weights repeatedly throughout the network. The later attention layer operates on contextualized feature representation generated by the previous layer. Learned linear (fully connected) transformations are applied to queries, keys, and values before attention computation. The patent discloses feed-forward subnetworks using activation function (e.g., ReLU on) together with the attention architecture.)
The method further comprises calculating second attention scores of pre-defined indicators based on the shared global map features, wherein the pre-defined indicators are processed by a second fully connected layer (In Col. 4, lines 15–29, Shazeer discloses that the weighted vectors are combined (weighted sum) to generate a new context representation that is forwarded to later layers.)
The second attention scores are determined based on activation functions, wherein map task- specific weighted representations are calculated based on the fusion of the second attention scores with the shared global map features of the second sub-neural network (In Col. 5, lines 35–57, Shazeer discloses using an activation function (e.g., ReLU) within the neural network feed-forward layer. The encoder subnetwork may also include a connection layer.)
Task-specific fused representations are determined based on the map task-specific weighted representations and the trajectory task-specific weighted representations (In Col.8, lines 38–50, Shazeer discloses later that attention sublayers compute another set of attention weights during a subsequent attention operation. Encoder subnetworks generate contextual encoded representation shared throughout the network. The attention mechanism computes attention weights and applies those weights to the encoded representation, producing a weighted output representation.)
Accordingly, it would have been obvious to a person of ordinary skill in the art, before the effective filing date of the claimed invention, having the teaching of Songtao in view of Etemad to include Shazeer, with an attention mechanism obtain the weights on the values. The attention sub-layer then computes a weighted sum of the values in accordance with these weights as taught by Shazeer. The motivation for doing so would have been to improve performance over conventional machine translation neural networks without task-specific tuning through the use of the attention mechanism (See (Col. 2, lines 28-30) of Shazeer.)
Regarding claim 17, Songtao in view of Etemad disclose elements of claim 15. Songtao in view of Etemad and Pack do not disclose:
The data processing system of claim 15, wherein the neural network further comprises a fully connected layer and wherein the neural network is further configured to determine attention scores of pre-defined indicators corresponding to road attributes based on the trajectory data, wherein the pre-defined indicators are processed by the fully connected layer, and wherein the attention scores are determined based on activation functions
However, Shazeer disclose the limitation (In Col. 5, lines 35–57, Shazeer discloses using an activation function (e.g., Rule) within the neural network feed-forward layer. The encoder subnetwork may also include a connection layer.)
Accordingly, it would have been obvious to a person of ordinary skill in the art, before the effective filing date of the claimed invention, having the teaching of Songtao in view of Etemad to include Shazeer, with an attention mechanism obtain the weights on the values. The attention sub-layer then computes a weighted sum of the values in accordance with these weights as taught by Shazeer. The motivation for doing so would have been to improve performance over conventional machine translation neural networks without task-specific tuning through the use of the attention mechanism (See (Col. 2, lines 28-30) of Shazeer.)
Regarding claim 19, Songtao in view of Etemad disclose elements of claim 18. Songtao in view of Etemad do not disclose:
The data processing system of claim 18, wherein the neural network is configured to fuse the attention scores with the shared global trajectory features of the first sub-neural network thereby generating trajectory task-specific weighted representations
However, Shazeer disclose the limitation (In Col.8, lines 38–50, Shazeer discloses later that attention sublayers compute another set of attention weights during a subsequent attention operation. Encoder subnetworks generate contextual encoded representation shared throughout the network. The attention mechanism computes attention weights and applies those weights to the encoded representation, producing a weighted output representation.)
Accordingly, it would have been obvious to a person of ordinary skill in the art, before the effective filing date of the claimed invention, having the teaching of Songtao in view of Etemad to include Shazeer, with an attention mechanism obtain the weights on the values. The attention sub-layer then computes a weighted sum of the values in accordance with these weights as taught by Shazeer. The motivation for doing so would have been to improve performance over conventional machine translation neural networks without task-specific tuning through the use of the attention mechanism (See (Col. 2, lines 28-30) of Shazeer.)
Claim 26 is rejected under 35 U.S.C. 103 as being unpatentable over “Fusing Taxi Trajectories and RS Images to Build Road Map via DCNN”, https://ieeexplore.ieee.org/stamp/stamp.jsp?arnumber=8891700, Li et al, 11/5//2019, hereinafter referred to as Li, in view of “Neural Network Based Uncertainty Prediction for Autonomous Vehicle Application
”,https://www.frontiersin.org/journals/neurorobotics/articles/10.3389/fnbot.2019.00012/full, Zhang et al, 05/10/2019, hereinafter referred to as Zhang and further in view of “CAR-Net: Clairvoyant Attentive Recurrent Network”, https://arxiv.org/abs/1711.10061, Amir et al, 2018, hereinafter referred to as Amir.
With respect to claim 26, Li discloses:
A method for training an automated predictor, the method comprising: performing forward propagation by inputting training data into the automated predictor to obtain an output result, for a plurality of road segments of a geographical area, wherein the training data comprises (On page 4 (IV. ROAD IMAGE LEARNING, last paragraph), Li teaches inputting training data comprising trajectory feature graphs and remote-sensing image feature maps into a deep convolutional neural network (RUNET), wherein the network processes the input feature to generate an output comprising road/non-road classification maps. Thus, Li teaches or suggests forward processing/propagation of the training data through the neural network to obtain an output result for road portions of a geographical area.)
Trajectory features (On page 2 (A. EXTRACTING ROAD FROM RS IMAGESBASED ON DCNN, B. EXTRACTING ROAD FROM TRAJECTORIES and C. EXTRACTING ROAD BY COMBINING RS IMAGESWITH TRAJECTORIES), Li teaches a taxi GPS trajectory data and explains that trajectories contain information about both the geometry of the road network and the movement pattern of vehicles. It then extracts multiple features to construct what it calls trajectory-based road feature graphs (T-RF).)
The trajectory features extracted from trajectory data received as raw trace data recorded by one or more vehicles traversing at least a portion of the geographical area and the map features extracted from map data (On page 2, Li teaches taxi GPS trajectory data and extracts multiple road-related characteristics. It states that "feature graphs with trajectory density and speed are obtained from taxi GPS trajectory data." More specifically, Li extracts an original grid feature, trajectory-density feature, and vehicle-speed feature, then integrates these into a trajectory-based road feature graph (T-RF).)
A classifier configured to classify an output of the neural network into prediction probabilities of the road attributes (On page 5 (paragraph 1-2), Li teaches a neural-network classifier configured to classify neural-network output into road-related classes, wherein the output of the training model comprises two-category classification maps identifying road and non-road portions. Li further employs a sigmoid output layer for classification. Thus, Li teaches the underlying neural network classification architecture for producing road-related classification outputs.)
Wherein the neural network comprises a first sub-neural network for the trajectory features and a second sub-neural network for the map features (On page 3 (second paragraph), Li teaches taxi GPS trajectory data and explains that trajectory data reveal both road-network geometry and vehicle movement patterns. Li extracts multiple features: the original grid feature represents road geometry/location based on trajectory points. Trajectory-density feature — represents the density/distribution of GPS trajectory points. Speed feature — represents vehicle movement patterns by calculating the average speed.)
With respect to claim 26, Li do not explicitly disclose:
Performing back propagation according to a difference between the output result and an expected result to adjust weights of the automated predictor
Repeating the above steps until a pre-determined convergence threshold is achieved, wherein the automated predictor comprises
A neural network configured to predict road attributes by generating an output of task-specific fused features using attention scores applied to trajectory features and map features,
However, it is known by Zhang to disclose:
Performing back propagation according to a difference between the output result and an expected result to adjust weights of the automated predictor (On pages 9-10 (5.1 Error Measurement & 5.2 Training Algorithm), Zhang teaches training a neural network using an Error Backpropagation (EBP) technique, wherein network performance is evaluated based on error measurement between the neural-network output and target/reference data. Zhang further teaches iterative neural-network training using the Levenberg-Marquardt algorithm, which employs Error Back propagation depending upon the error characteristics during convergence, thereby teaching or suggesting back propagating an output error for adjusting the neural network during training.)
Repeating the above steps until a pre-determined convergence threshold is achieved, wherein the automated predictor comprises (On pages 9-10 (5.1 Error Measurement), Zhang teaches that the neural networks are trained until either a certain performance value or a specified number of iteration/epochs is reached, and that early stopping is used. It also discusses convergence of the training algorithm and expressly identified Error Back propagation (EBP).)
Li and Zhang are analogous pieces of art because both concern neural-network processing of vehicle trajectory/road-related data. Accordingly, it would have been obvious to a person of ordinary skill in the art, before the effective filing date of the claimed invention, to modify Li, with extracting road-related characteristics from trajectory data and image data, provides those features to a deep neural network as taught by Li, with determining the location uncertainty in the trajectory as taught by Zhang. The motivation for doing so would have been to improve accuracy continuity, and completeness of automatically generated road maps (See (Page 2, paragraph 1) of Li.)
With respect to claim 26, Li in view of Zhang do not disclose:
A neural network configured to predict road attributes by generating an output of task-specific fused features using attention scores applied to trajectory features and map features
However, it is known by Amir to disclose:
A neural network configured to predict road attributes by generating an output of task-specific fused features using attention scores applied to trajectory features and map features (On page 4 (3 CAR-Net), Amir teaches an attentive neural-network architecture that jointly utilizes trajectory information and electronic image information representing a navigation environment. The network receives a past motion trajectory and a top-view image of the navigation scene and applies an attention mechanism to identify image region relevant to the prediction task, including road intersections. Accordingly, Amir teaches applying learned attention values to features derived from trajectory/navigation information and electronic image information for generating a task-relevant representation used by a neural predictor.)
Li in view of Zhang and Amir are analogous pieces of art because both concern neural-network processing of vehicle trajectory/road-related data. Accordingly, it would have been obvious to a person of ordinary skill in the art, before the effective filing date of the claimed invention, to modify Amir, with multi-source attention, the mechanism applies weights to spatial areas of the scene based on their importance as taught by Amir. The motivation for doing so would have been to improve trajectory/path prediction, particularly by improving how a neural network uses the surrounding scene or map information when predicting where an agent or vehicle will move next (See (Page 4, 4.3 Ablation Study) of Amir.)
Response to Arguments
Applicant's arguments filed on 05/18/2026 have been fully considered, and in part are persuasive.
Pertaining to Rejection under 101
Rejections for claims 1-19 and 26 are withdrawn under 35 USC 101.
Pertaining to Rejection under 103
Applicant’s arguments in regard to the examiner’s rejections under 35 USC 103 are moot in view of the new grounds of rejection.
Conclusion
Any inquiry concerning this communication or earlier communications from the examiner should be directed to EVEL HONORE whose telephone number is (703)756-1179. The examiner can normally be reached Monday-Friday 8 a.m. -5:30 p.m.
Examiner interviews are available via telephone, in-person, and video conferencing using a USPTO supplied web-based collaboration tool. To schedule an interview, applicant is encouraged to use the USPTO Automated Interview Request (AIR) at http://www.uspto.gov/interviewpractice.
If attempts to reach the examiner by telephone are unsuccessful, the examiner’s supervisor, Mariela D Reyes can be reached at (571) 270-1006. The fax phone number for the organization where this application or proceeding is assigned is 571-273-8300.
Information regarding the status of published or unpublished applications may be obtained from Patent Center. Unpublished application information in Patent Center is available to registered users. To file and manage patent submissions in Patent Center, visit: https://patentcenter.uspto.gov. Visit https://www.uspto.gov/patents/apply/patent-center for more information about Patent Center and https://www.uspto.gov/patents/docx for information about filing in DOCX format. For additional questions, contact the Electronic Business Center (EBC) at 866-217-9197 (toll-free). If you would like assistance from a USPTO Customer Service Representative, call 800-786-9199 (IN USA OR CANADA) or 571-272-1000.
EVEL HONORE
Examiner
Art Unit 2142
/HAIMEI JIANG/Primary Examiner, Art Unit 2142
Prosecution Insights