Track refinement networks
Provided are methods for a track refinement network. In examples, center boxes are obtained from a record of driving data, wherein a center box is a center of a sequence of boxes along a track, and the track is associated with a tracked object detected within the sequence of boxes, each respective box comprising a center, a size, and an orientation. Track windows are generated around respective center boxes, wherein a track window corresponds to a respective center box along the track. Track windows are cropped and normalized with respect to center boxes to enable single refinement model for multiple object classes. Point cloud features and trajectory features are extracted from the cropped and normalized track windows. The point cloud features and trajectory features are input into a track refinement network, wherein the track refinement network uses features from the entire track to output a refined center, a refined size, and a refined orientation of each respective center box.
1 . A system, comprising:
at least one processor, and
at least one non-transitory storage media storing instructions that, when executed by the at least one processor, cause the at least one processor to:
obtain center boxes from a record of driving data, wherein the center boxes form a sequence of boxes along a track, and the track is associated with a tracked object detected within the sequence of boxes, each respective center box comprising a center, a size, and an orientation;
generate track windows around respective center boxes, wherein a track window corresponds to a respective center box along the track;
crop and normalize the track windows with respect to the center boxes;
extract point cloud features and trajectory features from the cropped and normalized track windows;
input the point cloud features and trajectory features into a track refinement network, wherein the track refinement network uses features from the track to output a refined center, a refined size, and a refined orientation of each respective center box; and
deploy the refined center, size, and orientation of each respective center box.
2 . The system of claim 1 , wherein the track refinement network regresses a residual between the respective center box obtained from an offline perception system and a ground truth box.
3 . The system of claim 1 , wherein normalization scales a canvas based on the respective center box.
4 . The system of claim 1 , wherein the track refinement network refines the center boxes corresponding to multiple classifications associated with object detection.
5 . The system of claim 1 , wherein the track refinement network comprises shared layers with multiple task heads, and the track refinement network enables execution of multiple tasks corresponding to the multiple task heads.
6 . The system of claim 1 , wherein the track refinement network further outputs track attributes.
7 . The system of claim 1 , wherein the track refinement network is trained using an auxiliary loss function that smooths output of a trained track refinement network.
8 . The system of claim 1 , wherein deploying driving data comprises automatically generating a database of auto labelled training data.
9 . The system of claim 1 , wherein deploying driving data comprises inputting center boxes to an online tracker in online perception.
10 . The system of claim 1 , wherein deploying driving data comprises generating refined boxes to enable image-LiDAR fusion.
11 . The system of claim 1 , comprising enforcing at least one constraint on the refined center, size, and orientation of each respective center box.
12 . A method, comprising:
obtaining, with at least one processor, center boxes from a record of driving data, wherein the center boxes form a sequence of boxes along a track, and the track is associated with a tracked object detected within the sequence of boxes, each respective center box comprising a center, a size, and an orientation;
generating, with the at least one processor, track windows around respective center boxes, wherein a track window corresponds to a respective center box along the track;
cropping and normalizing, with the at least one processor, the track windows with respect to the center boxes;
extracting, with the at least one processor, point cloud features and trajectory features from the cropped and normalized track windows;
inputting, with the at least one processor, the point cloud features and trajectory features into a track refinement network, wherein the track refinement network uses features from the track to output a refined center, a refined size, and a refined orientation of each respective center box; and
deploying, with the at least one processor, the refined center, size, and orientation of each respective center box.
13 . The method of claim 12 , wherein the track refinement network regresses a residual between the respective center box obtained from an offline perception system and a ground truth box.
14 . The method of claim 12 , wherein normalization scales a canvas based on the respective center box.
15 . The method of claim 12 , wherein the track refinement network refines the center boxes corresponding to multiple classifications associated with object detection.
16 . The method of claim 12 , wherein the track refinement network comprises shared layers with multiple task heads, and the track refinement network enables execution of multiple tasks corresponding to the multiple task heads.
17 . At least one non-transitory storage media storing instructions that, when executed by at least one processor, cause the at least one processor to:
obtain center boxes from a record of driving data, wherein the center boxes form a sequence of boxes along a track, and the track is associated with a tracked object detected within the sequence of boxes, each respective center box comprising a center, a size, and an orientation;
generate track windows around respective center boxes, wherein a track window corresponds to a respective center box along the track;
crop and normalize the track windows with respect to the center boxes;
extract point cloud features and trajectory features from the cropped and normalized track windows;
input the point cloud features and trajectory features into a track refinement network, wherein the track refinement network uses features from the track to output a refined center, a refined size, and a refined orientation of each respective center box; and
deploy the refined center, size, and orientation of each respective center box.
18 . The at least one non-transitory storage media of claim 17 , wherein the track refinement network regresses a residual between the respective center box obtained from an offline perception system and a ground truth box.
19 . The at least one non-transitory storage media of claim 17 , wherein normalization scales a canvas based on the respective center box.
20 . The at least one non-transitory storage media of claim 17 , wherein the track refinement network refines the center boxes corresponding to multiple classifications associated with object detection.