基于改进YOLOv4的前方车辆检测方法

Forward vehicle detection method based on improved YOLOv4

针对现有深度学习算法对车辆检测精度不高、实时性不足以及对小目标检测能力较弱的问题,提出一种基于改进YOLOv4的前方车辆检测方法。针对小目标漏检率高的问题,使用密集连接块替代CSPResNet中的残差块,重复利用特征信息,提高小目标的检测能力;针对检测精度和实时性要求,将密集连接块结构修改为卷积、批归一化和激活操作,对结构中的卷积层和批归一化层进行融合并使用Mish激活函数,提高模型的检测精度和速度。实验结果表明,改进YOLOv4算法较原有算法参数量减少23%,检测速度和精度分别提升3.5fps和2.73%,同时提高了对小目标的检测能力。

Aiming at the problems that detection precision is not high,real-time performance is insufficient and detection ability of small targets is weak,that the existing deep learning algorithms has for vehicle detection,a forward vehicle detection method based on improved YOLOv4 is proposed.Aiming at the problem of high false negative rate of small targets,dense connection block(Dense Block)are used to replace the residual blocks in CSPResNet to reuse feature information and improve the detection ability of small targets.Aiming for the requirements of detection precision and real-time,the dense connection block structure is modified to convolution,batch normalization and activation operations,and the convolutional layer and batch normalization layer in the structure are fused and the Mish activation function is used to improve the detection precision and speed of the model.Experimental results show that the improved YOLOv4 algorithm reduces the number of parameters by 23% compared with the original algorithm,and the detection speed and precision are increased by 3.5 fps and 2.73%,respectively,and the detection ability of small targets is improved.