Methods and mobility apparatus for predicting agent behavior for autonomous driving
A method for controlling autonomous driving of a vehicle is introduced. The method involves generating, based on an image obtained from a sensor of a mobility device, a feature map and a feature of an object, generating an occupancy flow map feature of the object, wherein the occupancy flow map feature comprises local information associated with a predicted trajectory of the object in the feature map, wherein the feature of the object is fused with the feature map, and wherein the predicted trajectory is estimated from the feature of the object fused with the feature map, outputting a predicted path of the object from a feature of the object fused with the occupancy flow map feature, wherein the predicted trajectory is determined, based on a reliability of the predicted trajectory, as the predicted path, and controlling, based on the predicted path, autonomous driving of the vehicle.
1 . A method performed by a processor for controlling autonomous driving of a vehicle, the method comprising:
generating, based on an image obtained from a sensor of a mobility device, a feature map and a feature of an object, wherein the feature map is associated with a transformed view and a transformed viewpoint;
generating an occupancy flow map feature of the object, wherein the occupancy flow map feature comprises local information associated with a predicted trajectory of the object in the feature map that reflects the predicted trajectory, wherein the feature of the object is fused with the feature map, and wherein the predicted trajectory is estimated from the feature of the object fused with the feature map;
outputting a predicted path of the object from a feature of the object fused with the occupancy flow map feature, based on information, associated with a behavioral context of the object, derived from the image, wherein the predicted trajectory is determined, based on a reliability of the predicted trajectory, as the predicted path; and
controlling, based on the predicted path, autonomous driving of the vehicle.
2 . The method of claim 1 , wherein the generating the occupancy flow map feature comprises:
fusing the feature of the object with the feature map by cross attention to represent the predicted trajectory of the object in the feature map, wherein the cross attention uses the feature map with the predicted trajectory and the feature of the object;
extracting, based on the predicted trajectory from the feature map, a local area;
generating local information associated with the local area; and
providing the occupancy flow map feature comprising:
the feature map fused with the feature of the object, and
the local information.
3 . The method of claim 2 , wherein the generating the local information comprises:
extracting occupancy areas occupied by the predicted trajectory from the feature map;
predicting, based on the feature map, a trajectory of the object by identifying global interaction between the occupancy areas through self attention; and
generating the local information by applying a convolution neural network (CNN) to the feature map for which the self attention is processed, wherein the local information comprises a feature of the local area associated with the predicted trajectory.
4 . The method of claim 1 , wherein the feature of the object fused with the occupancy flow map feature is generated by deformable cross attention that uses the occupancy flow map feature and a feature of the object.
5 . The method of claim 1 , wherein the feature of the object fused with the feature map of the transformed view is a feature of the object that is processed to reflect interaction of the object.
6 . The method of claim 1 , further comprising, before the outputting the predicted path of the object;
generating, based on vision information representing a situation in the image, text data by using an image language transform model, wherein the text data describes the predicted trajectory of the object and a situation of the object in text;
encoding the text data; and
generating information of the object that represents the behavioral context, wherein the behavioral context comprises the predicted trajectory and the situation of the object in text that matches the feature of the object, and wherein the situation of the object in text is matched with the feature of the object based on a contrastive learning model that uses the encoded text data and the feature of the object as processed in a common embedding space.
7 . The method of claim 6 , wherein the generating the text data comprises:
refining, by a language refinement model and based on the vision information, the text data.
8 . The method of claim 6 , wherein a learning model for implementing the method is trained by using a loss function that comprises:
a loss based on a similarity between the text data, wherein the text data describes the situation of the object in text and a situation feature of the object,
a loss based on a trajectory of the object in the feature map, wherein the feature map is fused with the feature of the object, and
a loss based on the predicted trajectory of the object.
9 . The method of claim 1 , wherein the transformed view is a bird's eye view, and wherein:
the feature of the object is generated based on object information that has multi-modal data obtained from the image,
the multi-modal data is heterogeneous data, and
the feature of the object is generated by a transformer model and based on concatenated data that is combined by learning the multi-modal data.
10 . The method of claim 9 , wherein the multi-modal data comprises:
a time-series trajectory of the object,
a type of the object, and
temporal change data of the object, wherein the temporal change data comprises a distance and a heading displacement from a perspective of the mobility device.
11 . An apparatus for controlling autonomous driving of a vehicle, the apparatus comprising:
a sensor configured to detect a surrounding environment of a mobility device;
a memory configured to store at least one instruction; and
a processor configured to execute the at least one instruction stored in the memory,
wherein the processor is further configured to:
generate, based on an image obtained from the sensor, a feature map and a feature of an object, wherein the feature map is associated with a transformed view and a transformed viewpoint;
generate an occupancy flow map feature of the object, wherein the occupancy flow map feature comprises local information associated with a predicted trajectory of the object in the feature map that reflects the predicted trajectory, wherein the feature of the object is fused with the feature map, and wherein the predicted trajectory is estimated from the feature of the object fused with the feature map; and
output a predicted path of the object from a feature of the object fused with the occupancy flow map feature, based on information, associated with a behavioral context of the object derived from the image, wherein the predicted trajectory is determined, based on a reliability of the predicted trajectory, as the predicted path; and
control, based on the predicted path, autonomous driving of the vehicle.
12 . The apparatus of claim 11 , wherein the processor is further configured to generate the occupancy flow map feature by:
fusing the feature of the object with the feature map by cross attention represent the predicted trajectory of the object in the feature map, wherein the cross attention uses the feature map with the predicted trajectory and the feature of the object;
extracting, based on the predicted trajectory from the feature map, a local area;
generating local information associated with the local area; and
providing the occupancy flow map feature comprising:
the feature map fused with the feature of the object, and
the local information.
13 . The apparatus of claim 12 , wherein the processor is further configured to generate the local information by:
extracting occupancy areas occupied by the predicted trajectory from the feature map;
predicting, based on the feature map, a trajectory of the object by identifying global interaction between the occupancy areas through self attention; and
generating the local information by applying a convolution neural network (CNN) to the feature map for which the self attention is processed, wherein the local information comprises a feature of the local area associated with the predicted trajectory.
14 . The apparatus of claim 11 , wherein the feature of the object fused with the occupancy flow map feature is generated by deformable cross attention that uses the occupancy flow map feature and the feature of the object.
15 . The apparatus of claim 11 , wherein the feature of the object fused with the feature map of the transformed view is a feature of the object that is processed to reflect interaction of the object.
16 . The apparatus of claim 11 , wherein the processor is further configured to, before outputting the predicted path of the object:
generate, based on vision information representing a situation in the image, text data by using an image language transform model, wherein the text data describes the predicted trajectory of the object and a situation of the object in text;
encode the text data; and
generate information of the object that represents the behavioral context, wherein the behavioral context comprises the predicted trajectory and the situation of the object in text that matches the feature of the object, and wherein the situation of the object in text is matched with the feature of the object based on a contrastive learning model that uses the encoded text data and the feature of the object as processed in a common embedding space.
17 . The apparatus of claim 16 , wherein the processor is further configured to generate the text data by refining, by a language refinement model and based on the vision information, the text data.
18 . The apparatus of claim 16 , wherein a learning model is trained by using a loss function that comprises:
a loss based on a similarity between the text data, wherein the text data describes the situation of the object in text and a situation feature of the object,
a loss based on a trajectory of the object in the feature map, wherein the feature map is fused with the feature of the object, and
a loss based on the predicted trajectory of the object.
19 . The apparatus of claim 11 , wherein the transformed view is a bird's eye view, and wherein:
the feature of the object is generated based on object information that has multi-modal data obtained from the image,
the multi-modal data is heterogeneous data, and
the feature of the object is generated by a transformer model and based on concatenated data that is combined by learning the multi-modal data.
20 . The apparatus of claim 19 , wherein the multi-modal data comprises:
a time-series trajectory of the object,
a type of the object, and
temporal change data of the object, wherein the temporal change data comprises a distance and a heading displacement from a perspective of the mobility device.