基于YOLOv3的车辆行人目标检测算法改进方法研究  

RESEARCH ON IMPROVEMENT METHOD OF VEHICLE PEDESTRIAN TARGET DETECTION ALGORITHM BASED ON YOLOV3

作　　者：梁策[1] 王景中[1] 王宝成[1] Liang Ce;Wang Jingzhong;Wang Baocheng(School of Information Science and Technology,North China University of Technology,Beijing 100144,China)

机构地区：[1]北方工业大学信息学院,北京100144

出　　处：《计算机应用与软件》2022年第7期201-206,226,共7页Computer Applications and Software

摘　　要：针对车辆行人检测中目标尺度多、尺度小、目标遮挡严重的问题,将实时性与准确性较高的YOLOv3算法应用于行人车辆检测领域,并提出两点优化方法:(1)优化非大值抑制算法,高斯函数衰减的形式对预测框置信得分进行抑制,避免被遮挡目标预测框被误删,提升算法对被遮挡目标的检测能力;(2)优化YOLOv3的网络结构,增加更多的特征层与残差单元,获得更高分辨率、更多预测尺度的网络结构,提升对小目标及多尺度目标的识别能力。实验使用UA-DETRAC、PASCAL VOC数据集进行训练与测试,结果表明相较于传统YOLOv3算法,在行人、车辆目标尺度多的情况下,召回率有一定提高;在目标图像小及被遮挡的情况下,准确率有所提升。Aimed at the problem that vehicle pedestrian detection faces the test of multiple target scales, small scales, and severe target occlusion, the YOLOv3 algorithm with high real-time and high accuracy was applied to the field of pedestrian vehicle detection. Two optimization methods were proposed for the above problems:(1) the non-large value suppression algorithm was optimized, and the Gaussian function attenuation form was used to suppress the prediction frame confidence score, avoiding the occlusion target prediction frame being deleted by mistake, and improving the detection ability of the algorithm to the occluded target.(2) The network structure of YOLOv3 was optimized to add more feature layers and residual units, obtain higher-resolution and more predictive scale network structure, and improve the ability to identify small targets and multi-scale targets. The experiment used the UA-DETRAC and PASCAL VOC data sets for training and testing. The experimental results show that compared with the traditional YOLOv3 algorithm, the recall rate is improved in the case of pedestrians and vehicle targets. The accuracy rate is improved when the target image is small and occluded.

关 键 词：YOLOv3算法 人车识别 多尺度预测 

分 类 号：TP3[自动化与计算机技术—计算机科学与技术]