3-D Vehicle Detection Enhancement Using Tracking Feedback in Sparse Point Clouds Environments

Abstract

In recent years, vehicle detection in intelligent transportation systems using 3D LIDAR point clouds based on deep neural networks has made substantial progress. However, when the point clouds are very sparse, the detection model cannot generate proposals efficiently, resulting in false negative results. Considering that the object tracking technology accurately predicts vehicles based on historical measurements and motion models, and these prediction results can become proposals for object detection. Therefore, this paper proposes a novel object detection paradigm based on tracking feedback to address the false negative problem based on sparse point clouds. According to the distribution of the state vector from the Kalman prediction, multiple proposals are sampled and fed back to the second stage of two-stage detection models. After regression and non-maximum suppression, the false negative results can be effectively reduced. This method enhances the vehicle detection capability of classical neural networks. Comparing the recall metric of multiple detection models at different distances in the public KITTI and nuSences datasets, the proposed method can promote up to 5.31% compared to the previous method, which reflects the effectiveness and versatility of the proposed method.