复杂场景下的车辆检测算法研究及其优化

Research and Optimization of Vehicle Detection Algorithms in Complex Scenarios

智能辅助驾驶要解决的关键问题之一是车辆检测。为了解决车辆检测问题,本文用MobileNetv2替代YOLOv5的主干网络来对YOLOv5实现轻量化优化,并利用知识蒸馏对轻量化后的YOLOv5进行模型压缩,以解决模型参数量大的问题。实验结果表明,S-MobileNetv2-YOLOv5模型的mAP为0.743,参数量为3.60M,说明模型的目标识别具有较好的效果,在保证精度的同时大幅度减少了参数量,适合部署在硬件设备上。

One of the key issues that intelligent assisted driving needs to solve is vehicle detection.In order to solve the problem of vehicle detection,this article replaces the backbone network of YOLOv5 with MobileNetv2 to achieve lightweight optimization of YOLOv5,and uses knowledge distillation to compress the lightweight YOLOv5 model to solve the problem of large model parameters.The experimental results show that the mAP of the S-MobileNetv2-YOLOv5 model is 0.743,with a parameter size of 3.60M,indicating that the model has good target recognition performance,significantly reducing the parameter size while ensuring accuracy,and is suitable for deployment on hardware devices.