基于YOLOv7道路交通热红外图像目标检测算法

Object Detection Algorithm of Road Traffic Thermal Infrared Image Based on YOLOv7

热红外图像具有分辨率低、高噪声、空间相关性强的特点。为解决道路交通红外目标检测算法存在的检测精度低、漏检和误检等问题,提出一种改进YOLOv7的算法。将主干网络原有的ELAN模块替换成ELAN-P模块,降低模型的参数量和计算量,使模型更加轻量化,增强对红外目标的提取能力;在主干网络和颈部网络引入CA注意力机制,将坐标信息嵌入到通道中,增强对模糊目标和密集目标的定位能力;将原有的CIoU损失函数替换成WIoU损失函数,提高对遮挡目标和弱小目标的检测精度。在中国热红外数据集CTIR上实验表明,改进算法相较于YOLOv7算法,参数量和计算量分别减少11.6百分点和19.5百分点,检测精度mAP值提高了3.1百分点,其中Car、Pedestrian、Cyclist、Bus和Truck五个类别的检测结果AP值分别提高了1.9百分点、1.9百分点、1.5百分点、4.9百分点和5.3百分点,检测性能有所提升。在公开数据集FLIR上进行泛化性对比实验,结果表明改进算法具有通用性。

Thermal infrared images have the characteristics of low resolution,high noise and strong spatial correlation.To solve the problems of low detection accuracy,missed detection and false detection in infrared target detection algorithms for road traffic,an improved YOLOv7 algorithm was proposed.The ELAN module of the backbone network was replaced by the ELAN-P module,which reduced the number of parameters and calculation of the model,made the model more lightweight,and enhanced the ability to extract infrared targets.The CA mechanism was introduced into the backbone network and neck network,and the coordinate information was embedded into the channel to enhance the localization ability of fuzzy targets and dense targets.The original CIoU loss function was replaced with the WIoU loss function to improve the detection accuracy of occluded targets and dim and small targets.Experiments on the Chinese thermal infrared dataset CTIR show that compared with the YOLOv7 algorithm,the improved algorithm reduces the parameter amount and calculation amount by 11.6 percentage points and 19.5 percentage points respectively,and improves the detection accuracy mAP value by 3.1 percentage points.Among them,the AP value of the five categories of Car,Pedestrian,Cyclist,Bus and Truck increased by 1.9 percentage points,1.9 percentage points,1.5 percentage points,4.9 percentage points and 5.3 percentage points respectively,indic.