面向无人机视频分析的车辆目标检测方法

提出一种将航拍车辆视频预处理与轻量化目标检测模型结合的级联方式.首先,针对无人机拍摄的车辆视频数据大量冗余的问题,在边缘设备设置一个两级过滤器,通过帧的像素级及结构性差异过滤大量冗余帧,从而大幅减少传输到后端的检测模型的帧数;其次,针对高精度目标检测模型时延高的问题,采用通道剪枝与层剪枝结合的方法压缩YOLOv3模型并部署在PC端,实现时延和精度的均衡.实验结果表明:两级过滤器能够有效过滤90%以上的冗余帧数,相较于原模型,压缩模型在精度仅下降2%左右的情况下,检测速度提高78.3%,达到36.9帧·s^-1.

Vehicle Target Detection Method for Unmanned Aerial Vehicle Video Analysis

A cascading method combining aerial vehicle video preprocessing with lightweight target detection model was proposed. Firstly, aiming at the problem of massive redundancy of vehicle video data shot by unmanned aerial vehicle, a two-stage filter was set in the edge device to filter a large number of redundant frames through the pixel level and structural difference of the frames, so as to greatly reduce the number of frames transmitted to the detection model at the back end. Secondly, to solve the problem of high delay of high-precision target detection model, a combination of channel pruning and layer pruning was used to compress the YOLOv3 model and deploy it on PC to achieve the balance of delay and precision. The experimental results show that the two-stage filter can effectively filter more than 90% of the redundant frames. Compared with the original model, the detection speed of the compression model is increased by 78.3%, reaching 36.9 frames per second, when the accuracy of the compression model is only decreased by about 2%.