基于改进YOLOv5的车辆目标检测研究

针对现有目标检测算法在自动驾驶等领域的车辆目标检测中存在检测精度不高，实时性和鲁棒性较差等问题，本文提出一种基于YOLOv5的车辆目标检测方法.本文在YOLOv5s网络模型框架下，引入一次性聚合（OSA）模块优化主干网络结构，提升网络特征提取能力；并采用非局部注意力机制进行特征增强；同时利用加权非极大值抑制方法实现检测框筛选.实验结果表明，在自制车辆检测数据集上，改进网络模型与原YOLOv5s模型相比，平均准确率均值（mAP）提升3%，不同目标类检测的平均准确率（AP）均得到提升，且检测速度满足实时性要求，对于密集车辆和不同光照条件下均能较好实现车辆目标检测.

Vehicle object detection based on improved YOLOv5 method

In view of the problems of low detection accuracy, poor real-time and robustness of existing target detection algorithms in vehicle target detection in autonomous driving fields, a vehicle target detection method based on YOLOv5 is proposed. With the framework of YOLOv5s network model, a one-shot aggregation (OSA) module is introduced to optimize the backbone network structure and improve the network feature extraction capability. Non-local attention mechanism is used for feature enhancement. At the same time, the weighted non--maximum suppression method is used to filter the detection frame. The experimental results show that compared with the original YOLOv5s model, the mAP of the improved network model is improved by 3%, and the AP of different target detection classes is improved, and the detection speed meets the real-time requirements. For dense vehicles and under different illumination conditions, vehicle target detection can be better achieved.