Non-uniform imaging object detection method based on NU-YOLO

LiDAR (Light Laser Detection and Ranging) is a potential technology that can acquire multi-dimensional and multi-wavelength imaging information. A non-uniform sampling strategy is implemented to enhance the imaging speed, but this system's data can't be efficiently processed and detected. In this paper, we introduce a novel method for object detection in non-uniform imaging, aiming to improve scene analysis and enhance the object detection effect. We use the KITTI dataset with a non-uniform sampling strategy to generate a simulation dataset and design and train NU-YOLO (Non-uniform YOLO) to detect the object in non-uniform images. We build a nonuniform LiDAR system to verify the object detection performance in real scenarios and the speed boost of the LiDAR system. The simulation and experimental results show that NU-YOLO can detect objects in non-uniform images with a higher mAP50 of 7.4% than YOLO (You Only Look Once) v8. By using a non-uniform sampling strategy, NU-YOLO performs 3% better on the mAP50 metric than the model trained using a uniform sampling dataset with the same counts of sampling points. Notably, the non-uniform imaging system increases in speed compared to its uniform sampling counterpart without compromising imaging quality in critical areas. This work opens new application prospects for more intelligent in LiDAR.