基于YOLOv5算法的红外图像行人检测研究

Research on pedestrian detection in infrared image based on YOLOv5 algorithm

目标检测是自动驾驶的重要前提,是与外界信息交互的重要环节。针对夜间远处行人检测识别精度低、漏检的问题,提出一种针对检测小尺寸行人的YOLOv5-p4的夜间行人识别模型。首先,通过增加更小目标的检测层,引入BiFPN特征融合机制,防止小目标被噪声淹没,使网络模型可以更聚焦于物体的细小特征;同时使用K-means先验框聚类出更小目标的锚框,并且使用了多尺度的数据增强方法,增加模型的鲁棒性。使用了MetaAcon-C激活函数与EIoU回归损失函数使模型收敛效果更好,提升了算法远距离行人的检测的准确率。最后在红外行人数据集FLIR上验证改进后的YOLOv5-p4模型对于行人的检测能力,实验结果表明该方法与传统方法相比,准确率从86.9%提升到90.3%,适合用于红外图像中的行人检测。

Target detection is an important prerequisite of automatic driving and an important link of interaction with external information. Aiming at the problem of low accuracy and missing detection of distant pedestrians at night, a nighttime pedestrian recognition model of YOLOv5-p4 for detecting small sized pedestrians is proposed in this paper. Firstly, by adding a detection layer of smaller targets and introducing a BiFPN feature fusion mechanism to prevent small targets from being drowned by noise, the network model can be more focused on the small features of the object. At the same time, K-means prior frames are used to cluster anchor frames of smaller targets, and multi-scale data enhancement method is used to increase the robustness of the model. MetaAcon-C activation function and EIoU regression loss function are used to improve the model convergence effect and improve the accuracy of long-distance pedestrian detection algorithm. Finally, the improved YOLOv5-p4 model for pedestrian detection is verified on the infrared pedestrian data set FLIR. The experimental results show that compared with the traditional methods, the accuracy of this method is improved from 86.9 % to 90.3 %,which is suitable for pedestrian detection in infrared images.