[Background introduction]
Unmanned systems often face the task of tracking and reconnaissance of specific targets, which requires the unmanned system to have the ability to identify and locate targets. At the same time, tasks generally occur over a continuous time span, and the same target exists If the target is repeatedly discovered by a single robot or multiple robots, target re-identification is also required to de-duplicate the same target. Comrades who have read my previous blog will know that I have done work on target detection, visible light camera and lidar data fusion, and pedestrian re-identification. Yes, those tasks are all paving the way for the target recognition and positioning tasks here. .
【Algorithm Verification Platform】
ZED binocular camera
Robosence lidar
NVIDIA GeForce RTX 3070
[Algorithm Deployment Platform]
ZED binocular camera
Robosence lidar
Intel Frost Canyon NUC (I7 version)
Songling Scout MINI chassis
【Software Technology Stack】
zed-ros-wrapper
rslidar_sdk
data_fusion
darknet_ros
object_tracker
[Implementation flow chart]
【Effect video】
I'll take another photo and upload it when I have time. . .
[Later improvements]
1. The target recognition part uses the more lightweight Yolo-Lite model, and the target detection rate on the NUC can reach 20FPS.
2. The target re-identification part uses the more lightweight Mobile-Net network, which greatly improves the classification rate.
3. The fusion of point clouds and images has been optimized, and the accuracy of point cloud reprojection has been greatly improved.