Dynamic object tracking method used for port autonomous-driving vehicle

Abstract

The invention discloses a dynamic object tracking method used for a port autonomous-driving vehicle. Firstly, continuous-frame point clouds are fused and merged through lidar installed at a truck headand GPS positioning information to obtain relatively dense point cloud perception data; then the fused point cloud data are voxelized to form a grid 3D graph, and randomness of a point cloud is reduced; then height data of voxel 3D information is used to eliminate the ground, a previous frame of fused data are used at the same time to obtain difference, most static obstacles are eliminated, and only dynamic object information is retained; then clustering is carried out on a remaining point cloud through a density-based clustering algorithm to obtain moving target information of a current frame; and finally, moving targets of adjacent frames are used to carry out a heuristic tracking algorithm to calculate a center, a size, speed, a movement direction, a life cycle and a historical trajectory of a moving target of the current frame, and a tracking obstacle list of the current frame is output. According to the method, the amount of data processing is smaller, a sensing area is wide, andcost is lower.

Description

technical field [0001] The invention relates to a dynamic target tracking method for port self-driving vehicles, belonging to the technical field of artificial intelligence. Background technique [0002] With the development of on-board sensors and automotive automation technology, autonomous driving tasks in specific scenarios can be realized. Optical imaging perception through monocular or multi-eye cameras, millimeter-wave radar perception based on the Doppler effect, and multi-line lidar based on active lasers are currently mainstream unmanned driving perception solutions. However, due to problems such as ambient light and imaging field of view, the camera can obtain high-definition road semantic information, but the amount of data processing is large, and the accuracy of ranging based on monocular or binocular vision decreases with distance. Although the millimeter-wave radar can use the Doppler effect to detect and track dynamic targets at high speeds, the false detec...

AI Technical Summary

Problems solved by technology

However, due to the large size and height of container trucks exceeding 3 meters, it is difficult to achieve full coverage of the body and its surroundings, especially the ground at short distances, by installing a single radar on the top of the vehicle

At the same time, the cost of lidar with high wire bundle is higher, usually dozens of times that of a single 16-line radar

The target detection algorithm based on deep learning also relies heavily on the obstacle features of the dense point cloud obtained by dense laser beams, and it is difficult to obtain stable targets for point cloud features with fewer laser lines; at the same time, high-power GPU devices cannot Deployment in the vehicle environment makes the deep learning solution not very practical at present

Images