基于改进YOLOv5s的轻量化车流量检测算法

Lightweight traffic flow detection algorithm based on improved YOLOv5s

由于小型嵌入式设备计算资源有限,很多算法无法在车流量检测中运行完成.针对此问题,提出了一种改进YOLOv5s的轻量化车流量检测算法,以满足模型对硬件的计算能力需求.该算法通过使用轻量化特征提取网络ShuffleNetV2代替YOLOv5s原主干网络,减小模型的计算复杂度;融入注意力机制模块以增强网络对车流量的表示,提高其检测精度;通过使用数据增强策略,扩增训练样本,提升模型的鲁棒性.采用改进YOLOv5s算法检测视频车辆目标,结合DeepSort跟踪完成车流量检测,在UA-DETRAC车流量检测数据集上进行了实验.结果表明:所提出的轻量级车流量检测网络在Jetson TX2上的检测准确率为94.96%,检测速度达到15帧/s,实现车流量的实时检测,模型参数量少,满足嵌入式设备的限制.

Traffic flow detection is generally accomplished by small embedded devices.Many algorithms cannot run in the embedded devices due to their limited computing resources.To address this problem,this paper proposes an improved YOLOv5s lightweight vehicle flow detection algorithm to meet the requirements of hardware computing capacity.The algorithm uses the lightweight feature extraction network ShuffleNetV2 instead of YOLOv5s backbone network to reduce the computational complexity of the model.The attention mechanism module is integrated to enhance the representation of traffic flow and improve its detection accuracy.The robustness of the model was improved by using data enhancement strategies to amplify training samples.The improved YOLOv5s algorithm was used to detect video vehicle targets,and the vehicle flow was detected by combining with DeepSort tracking.Experimental results on UA-DETRAC traffic flow detection dataset show that the proposed lightweight traffic flow detection algorithm achieves an accuracy of 94.96%on Jetson TX2,and the detection speed reached 15 Frames/s,which achieves real-time traffic flow detection with less model parameters and meeting the limitations of embedded devices.