基于YOLO和排斥力损失函数的行人检测方法

针对行人检测受人体姿态复杂、光照变化、遮挡严重等影响,导致检测效率和精度不高的问题,提出一种基于YOLO和排斥力损失函数的行人检测方法.首先,对YOLO模型进行改进,主要是设置合适的预选框以及采用较大尺度的特征图进行特征提取,从而提高其对小物体的检测性能;然后,对排斥力损失函数进行改进,使其符合行人检测的应用场景,为接下来的融合检测提供新的损失函数;最后,将改进的YOLO和排斥力损失函数结合起来,利用YOLO模型速度快的特点提高运行速度,并利用排斥力损失函解决行人遮挡问题.在多个行人检测数据集上的实验结果表明:与其他算法相比,能够更加快速准确地实现行人检测.

A pedestrian detection method combining YOLO and repulsion loss

A pedestrian detection method combining YOLO and repulsion loss is proposed to solve the problem that pedestrian detection is affected by complex human body postures, illumination changes and severe occlusion, resulting in low detection efficiency and accuracy. Firstly, the improved YOLO model aims mainly to set the appropriate pre-selection box and feature extraction with larger scale feature maps to improve the detection performance of small objects. Then, the repulsion loss is improved to match the application scenario of pedestrian detection that provides a new loss function for the next fusion detection. Finally, the improved YOLO and the repulsion loss function are combined to improve the running speed by using the fast speed of YOLO, and the loss function is used to solve the pedestrian occlusion. The experimental results of occlusion problems on multiple pedestrian detection datasets show that this pedestrian detection can be implemented more quickly and accurately than other algorithms.