改进YOLOv4的复杂交通场景目标检测方法

针对复杂交通场景下密集小目标居多、目标尺寸差异大、目标间遮挡严重的问题，提出了一种基于YOLOv4框架的复杂交通场景下的目标检测算法。首先，构造多尺度特征融合提取模块作为主干网络特征提取模块，充分提取不同尺度目标特征信息，同时引入轻量化Ghost模块对主干网络特征进行维度调整；其次，将卷积模块与自注意力机制融合，构造倒残差自注意力模块应用到主干网络深层，深层网络在充分提取局部特征信息基础上获得了全局感知；然后，构造轻量级混合注意力模块，抑制背景噪声，增强密集小目标检测能力；最后，在Udacity数据集上进行实验，检测精度达到了84.41%,相比较YOLOv4, mAP(mean average precision)提高了3.07%,对1 920×1 200分辨率图像的检测FPS(frames per second)可达到49,提高了22.5%,精度提升的前提下实现了较好的实时性，更适用于复杂交通场景下的目标检测任务。

Improved YOLOv4 target detection method for complex traffic scenes

Aiming at the problems of most dense small targets, large target size differences and serious inter-target occlusion in complex traffic scenes, a target detection algorithm based on YOLOv4 framework for complex traffic scenes is proposed. Firstly, the multi-scale feature fusion extraction module is constructed as the backbone network feature extraction module to fully extract the target feature information at different scales, while the light-weight Ghost module is introduced to dimensionally adjust the backbone network features; secondly, the convolution module is fused with the self-attention mechanism to construct the inverse residual self-attention module to be applied to the deeper layer of the backbone network, and the deeper network obtains global perception based on fully extracting the local feature information; then, the global perception is obtained; then, the lightweight hybrid attention module is constructed to suppress the background noise and enhance the dense small target detection; finally, experiments were conducted on the Udacity dataset, and the detection accuracy reached 84.41%, compared with YOLOv4, the mAP (mean average precision) was improved by 3.07%, and the detection FPS (frames per second) for 1920×1200 resolution images reached 49, which was improved by 22.5%. The accuracy is improved with better real-time performance, which is more suitable for target detection tasks in complex traffic scenes.