基于改进YOLOv5s的自动驾驶中运动目标检测方法

Moving objects detection method in automatic driving based on improved YOLOv5s

针对自动驾驶领域运动目标检测存在目标相互遮挡、对小目标漏检误检严重、实时性和鲁棒性较差等问题,提出一种基于改进YOLOv5s的方法。将全局注意力模块嵌入到主干网络C3结构的残差块中,增强了网络的整体特征提取能力;将原始网络20×20的检测尺度替换为160×160,提高模型对于小目标的识别能力;采用SIoU Loss作为边界框回归损失函数,使预测框的定位更加精准。实验结果表明,该改进算法在公开的自动驾驶数据集SODA10M上均值平均精度达到了81.3%,较原始网络提升了7.5%,推理速度达到238 frames/s,与主流的目标检测算法相比也表现出一定的优越性。

Aiming at the problems of moving objects detection in the field of automatic driving,such as mutual occlusion of objects,serious missing and false detection of small objects,and poor real-time performance and robustness,a method based on improved YOLOv5s was proposed.The global attention module is embedded into the residual block in the C3 structure of the backbone network to enhance the overall feature extraction capability of the network.The detection scale of the original network 20×20 is replaced by 160×160 to improve the recognition ability of the model for small objects.SIoU Loss was used as the bounding box regression loss function,so as to make the positioning of the prediction box more accurate.The experimental results show that the mean average precision of the improved algorithm on the open automatic driving dataset SODA10M is up to 81.3%,7.5%higher than the original network,and the reasoning speed is up to 238 frames/s,which also shows certain superiority over the mainstream object detection algorithm.