复杂场景轻量级SSD行人检测方法

A Lightweight SSD Pedestrian Detection Method Under Complex Scenes

在高密度复杂场景下,传统目标检测算法由于缺乏对场景辅助信息的考虑,导致检测精度及性能难以满足实际需求。针对这一问题,本文提出了一种基于场景先验的轻量级SSD行人检测方法,根据地铁复杂场景中的行人先验特征及分布特点,重新设计SSD网络结构,添加小目标检测层,并调整生成候选框参数。实验表明:该方法在地铁场景下的行人平均检测精度(mAP)0.5,单帧检测耗时为9ms,相比经典SSD算法的行人检测精度和速度分别提升了6.2和43%,验证了所提方法的有效性。

Without the consideration of scene auxiliary information,the traditional algorithms of target detection have bad performance under high density complex scene.In order to solve this problem,we propose a lightweight SSD pedestrian detection method based on scene prior.Firstly,we analyze the prior and distribution characteristics of pedestrians in complex subway scenes.Then we redesign the structure of SSD network with extra small target detection layer.Meanwhile,we adjust the parameters of default box.After that,the model is optimized by training with large samples.Finally,the effectiveness of the proposed algorithm is verified by actual data experiments.The mean average precision(mAP)of pedestrian under subway scene is 90.5,and the single frame detection time is 9ms.The respective improving pedestrian detection precision and speed compared with classical SSD algorithm is 6.2 and 43%.