A multi-target recognition and tracking method for road traffic scenes

Abstract

The invention discloses a multi-target identification and tracking method oriented to road traffic scenes. The method combines the characteristics of the road traffic scene environment and obstacles to detect the dynamic obstacles around the road scene through lidar. The steps are as follows: first, the lidar data is preprocessed and converted into a binary image, and then the background difference method is used. The dynamic obstacles are extracted and clustered, and then the PointCNN algorithm is used for classification and identification, and the pose transformation of the obstacles is obtained through the iterative nearest neighbor algorithm. Finally, the UKF is used to track the obstacles respectively. The multi-target identification and tracking method proposed by the invention has good real-time performance, identification accuracy and tracking accuracy, and can effectively realize the safety monitoring of road traffic scenes.

Description

technical field [0001] The invention relates to the field of road traffic safety, in particular to a multi-target identification and tracking method oriented to road traffic scenes. Background technique [0002] With the development of economy and society, the problem of road traffic safety has become increasingly prominent. In road traffic accidents and secondary accidents in the rescue process, a large proportion of cases are caused by blind spots, which seriously endanger the safety of life and property of the public. For the prevention of traffic accidents, roadside sensing units (such as 3D LiDAR) can be arranged to monitor the road conditions in real time to prevent the occurrence of traffic accidents, such as figure 2 For road rescue scenarios, on-board sensing units (such as 3D LiDAR) can be used to safely monitor road rescue scenarios, thereby reducing the possibility of secondary accidents, such as image 3 shown. [0003] At present, the environmental monitorin...

AI Technical Summary

Problems solved by technology

However, such traditional methods have the following problems: the traditional method of manually extracting features for object recognition and classification only extracts features from limited angles such as curvature and normal vectors, and cannot fully utilize all the information of the 3D point cloud, and these features are only useful for a certain Some specific transformations have invariance, poor generalization ability, and it is difficult to accurately obtain the category of obstacles; the lack of obstacle point clouds caused by occlusion, etc., makes the cluster centroids unable to accurately replace the obstacle centroids, resulting in reduced tracking accuracy

Therefore, this method cannot be applied to the above two road traffic scenarios

Images