Light-weight YOLO based object detection algorithm for unmanned aerial vehicle

Abstract

Object detection algorithms based on deep learning usually have good results in terms of speed and accuracy on GPU- based computing platforms. However, as this kind of algorithm is not perfectly supported for CPU-based Unmanned aerial vehicle(UAV), the object detection algorithm usually used in UAV has the problem of slow detection speed, which will lead to traffic accidents, traffic congestion, and other problems. To solve this problem, we proposed an object detection algorithm based on YOLOv5. Firstly, aiming at the problem of light- weight model architecture, mobilenetv3 was added to YOLOv5 to replace the original backbone. Secondly, in order to maintain a high detection accuracy, omni-dimensional dynamic convolution was added in the feature fusion part of the network as a replacement for stander convolution. Through the architecture analysis, the proposed algorithm solves the problem in the UAV traffic monitoring system.