Surround vehicle tracking and motion prediction
A surround multi-object tracking and surround vehicle motion prediction framework is provided. A full-surround camera array and LiDAR sensor based approach provides for multi-object tracking for autonomous vehicles. The multi-object tracking incorporates a fusion scheme to handle object proposals from the different sensors within the calibrated camera array. A motion prediction framework leverages the instantaneous motion of vehicles, an understanding of motion patterns of freeway traffic, and the effect of inter-vehicle interactions. The motion prediction framework incorporates probabilistic modeling of surround vehicle trajectories. Additionally, subcategorizing trajectories based on maneuver classes leads to better modeling of motion patterns. A model takes into account interactions between surround vehicles for simultaneously predicting each of their motion.
1. A computer-implemented method, comprising:
receiving, by at least one data processor and from one or more sensors positioned on a vehicle, data indicative of a detected object in an area at least partially surrounding the vehicle;
fusing, by the at least one data processor, object proposals of the detected object identified from the received data to form a fused object proposal such that the fused object proposal is representative of the detected object, wherein fusing object proposals comprises:
projecting, by the at least one data processor, the object proposals onto a global coordinate frame with respect to the vehicle; and
combining, by the at least one data processor and based on the projection of the object proposals, a subset of object proposals determined to belong to the detected object;
generating, by the processing device, a track of the detected object, wherein the track is based on the fused object proposal and updated received data indicative of the detected object;
determining, by the at least one data processor, a maneuver class of the detected object, the maneuver class indicative of motion of the detected object and based on the generated track;
determining, by the at least one data processor, a plurality of future trajectories for the detected object, the plurality of future trajectories based on the generated track and at least one of a motion model or a probabilistic model; and
determining, by the processing device, an expected future trajectory of the detected object based on the corresponding maneuver class and the plurality of future trajectories.
2. The computer-implemented method of claim 1 , wherein the received data comprises:
a video feed, from a plurality of cameras positioned on the vehicle, of the area at least partially surrounding the vehicle; and
at least one point cloud, from at least one LiDAR sensor positioned on the vehicle, the at least one point cloud comprising LiDAR data associated with the area at least partially surrounding the vehicle.
3. The computer-implemented method of claim 1 , wherein generating the track of the detected object comprises:
identifying, by the at least one data processor, a status of the detected object; and
identifying, by the at least one data processor and if the status is an active status, a sequence of positions of the detected object based on the updated received data.
4. The computer-implemented method of claim 3 , wherein the status of the detected object is identified by a Markov Decision Process.
5. The computer-implemented method of claim 1 , wherein the maneuver class of the detected object is based on a Hidden Markov Model applied to a period of time of the generated track.
6. The computer-implemented method of claim 1 , wherein determining the plurality of future trajectories for the detected object comprises:
averaging, by the at least one data processor, predicted future locations and covariances provided by the probabilistic model.
7. The computer-implemented method of claim 1 , wherein the motion model comprises a constant velocity motion model, a constant acceleration motion model, and a constant turn rate and velocity motion model.
8. The computer-implemented method of claim 1 , wherein the expected future trajectory for the detected object is based on a feasibility assessment of each of the plurality of future trajectories, wherein the feasibility assessment is based on a configuration of other detected objects within a predefined range of the vehicle.
9. The computer-implemented method of claim 1 , further comprising:
generating, by the at least one data processor, an output representative of the expected future trajectory of the detected object.
10. A system, comprising:
at least one data processor; and
at least one memory storing instructions which, when executed by the at least one data processor, result in operations comprising:
receiving, from one or more sensors positioned on a vehicle, data indicative of a detected object in an area at least partially surrounding the vehicle;
fusing object proposals of the detected object identified from the received data to form a fused object proposal such that the fused object proposal is representative of the detected object, wherein fusing object proposals comprises:
projecting the object proposals onto a global coordinate frame with respect to the vehicle; and
combining, based on the projection of the object proposals, a subset of object proposals determined to belong to the detected object;
generating a track of the detected object, wherein the track is based on the fused object proposal and updated received data indicative of the detected object;
determining a maneuver class of the detected object, the maneuver class indicative of motion of the detected object and based on the generated track;
determining a plurality of future trajectories for the detected object, the plurality of future trajectories based on the generated track and at least one of a motion model or a probabilistic model; and
determining an expected future trajectory of the detected object based on the corresponding maneuver class and the plurality of future trajectories.
11. The system of claim 10 , wherein the received data comprises:
a video feed, from a plurality of cameras positioned on the vehicle, of the area at least partially surrounding the vehicle; and
at least one point cloud, from at least one LiDAR sensor positioned on the vehicle, the at least one point cloud comprising LiDAR data associated with the area at least partially surrounding the vehicle.
12. The system of claim 10 , wherein generating the track of the detected object comprises:
identifying a status of the detected object; and
identifying, if the status is an active status, a sequence of positions of the detected object based on the updated received data.
13. The system of claim 12 , wherein the status of the detected object is identified by a Markov Decision Process.
14. The system of claim 10 , wherein the maneuver class of the detected object is based on a Hidden Markov Model applied to a period of time of the generated track.
15. The system of claim 10 , wherein determining the plurality of future trajectories for the detected object comprises:
averaging, by the at least one data processor, predicted future locations and covariances provided by the probabilistic model.
16. The system of claim 10 , wherein the motion model comprises a constant velocity motion model, a constant acceleration motion model, and a constant turn rate and velocity motion model.
17. The system of claim 10 , wherein the expected future trajectory for the detected object is based on a feasibility assessment of each of the plurality of future trajectories, wherein the feasibility assessment is based on a configuration of other detected objects within a predefined range of the vehicle.
18. The system of claim 10 , wherein execution of the at least one memory storing instructions further result in:
generating a control signal indicative of the expected future trajectory of the detected object.
19. A non-transitory computer-readable storage medium including program code, which when executed by at least one data processor, causes operations comprising:
receiving, by at least one data processor a processing and from one or more sensors positioned on a vehicle, data indicative of a detected object in an area at least partially surrounding the vehicle;
fusing, by the at least one data processor, object proposals of the detected object identified from the received data to form a fused object proposal such that the fused object proposal is representative of the detected object, wherein fusing object proposals comprises:
projecting, by the at least one data processor, the object proposals onto a global coordinate frame with respect to the vehicle; and
combining, by the at least one data processor and based on the projection of the object proposals, a subset of object proposals determined to belong to the detected object;
generating, by the at least one data processor, a track of the detected object, wherein the track is based on the fused object proposal and updated received data indicative of the detected object;
determining, by the at least one data processor, a maneuver class of the detected object, the maneuver class indicative of motion of the detected object and based on the generated track;
determining, by the at least one data processor, a plurality of future trajectories for the detected object, the plurality of future trajectories based on the generated track and at least one of a motion model or a probabilistic model; and
determining, by the at least one data processor, an expected future trajectory of the detected object based on the corresponding maneuver class and the plurality of future trajectories.