基于改进YOLOv5的车辆红外图像多目标识别方法

Vehicle Infrared Image Multi-target Recognition Based on Improved YOLOv5

摘要:

城乡结合部的建设是城市建设中重要的一环，由于难以布设有效的检测设备，该区域车辆目标的夜间监管一直是城市管理的难题。基于无人机平台红外夜视图像多运动目标检测为解决这一难题提供了智能化路径：一种基于改进YOLOv5的红外夜视条件下多运动目标识别方法，分析了交通对象特征、车辆停放对道路红外辐射影响等，引入了CBAM注意力机制，以提取和融合空间和通道信息，增强了网络对目标的表达能力；综合Efficient IOU Loss和Focal Loss的优点，使用EIoU-Focal Loss损失函数替换了CIoU Loss函数，解决了样本不平衡、红外图像的低分辨率、噪声干扰大以及目标与背景对比度低等弊端，提高了检测的准确性；通过加入DCN动态调整卷积核的形状，适应图像中目标的形变，降低因形状不规则、变化较多造成的识别影响。在公开数据集上对改进网络与经典网络进行实验和数据比较，结果表明：综合改进后的网络对于多目标的识别,在YOLOv5x网络较高的识别结果基础上,精度提升3.9%,召回率提升4.1%,F1增加4.4%。

Abstract:

The rural-urban fringe is an important part of urban construction. Due to the difficulty of deploying effective detection equipment, the night supervision of vehicle targets in this area has been a difficult problem for urban management. Multi-moving target detection based on infrared night vision images of UAV platform provides an intelligent path to solve this problem: A multi-moving target recognition method based on improved YOLOv5 under infrared night vision conditions analyzed the characteristics of traffic objects and the impact of vehicle parking on road infrared radiation, etc. CBAM attention mechanism was introduced to extract and integrate spatial and channel information to enhance the expression ability of the network to the target. Combining the advantages of Efficient IOU Loss and Focal Loss, the EIoU-Focal loss function was used to replace CIoU loss function, which solved the disadvantages of sample imbalance, low resolution of infrared image, large noise interference and low contrast between target and background, and improved the detection accuracy. By adding DCN to dynamically adjust the shape of the convolution kernel, it can adapt to the deformation of the object in the image, and reduce the recognition influence caused by irregular shape and many changes. Finally, experiments and data comparisons indicate that the improved network based on YOLOv5 achieves higher recognition results and accuracy.