基于改进YOLOv3的站口行人检测方法

针对YOLOv3算法在行人检测上准确率低和漏检率高的问题,提出一种改进型YOLOv3的行人检测方法,并将其定义为GA-Wide-YOLOv3。该方法首先以行人头肩小目标为检测对象,进行重构数据集,利用遗传算法重新对目标先验框进行聚类,优化anchor参数,提高先验框与数据集的重合程度;其次改进YOLOv3,通过加宽网络宽度、减少网络深度,获得针对小目标检测的较大视野阈,避免梯度消失;最后,将多尺度检测算法3个yolo层前的1*1,3*3的卷积组各去掉2组,减少头肩小目标在复杂背景下的漏检率。在收集的数据集HS6936上进行对比实验,结果表明,基于遗传算法改进的K-means算法,平均交并比为81.89%,提高了0.8%;改进的YOLOv3算法检测平均准确率(mAP)为75.35%,召回率为81.20%,查准率为99.99%,较原始YOLOv3算法分别提高了2.53%,0.88%和2.75%。

Pedestrian detection method for station based on improved YOLOv3

Aiming at the problem of low accuracy and high missing rate of pedestrian detection in yorov3 algorithm, an improved yorov3 pedestrian detection method was proposed, which was defined as GA-Wide-YOLOv3. Firstly, the small head and shoulder targets of pedestrians were used as the detection objects to reconstruct the data set. The priori frames of the targets were clustered again by genetic algorithm. The anchor parameters were optimized to improve the priori frames and the weight of the data set. Secondly, YOLOv3 was improved. By widening the width of the network and reducing the depth of the network, the larger visual field threshold for small target detection was obtained to avoid the disappearance of the gradient. Finally, the convolution groups of 1^*1 and 3^*3 in front of the three Yolo layers of the multi-scale detection algorithm were removed from two groups respectively to reduce the missed detection rate of head shoulder small target in complex background. The comparative experiment was carried out on the collected data set hs6936, and the results were summarized. The results show that the improved K-means algorithm based on genetic algorithm has an average intersection and union ratio of 81.89%, which is 0.8% higher. The improved YOLOv3 algorithm has an average detection accuracy of 75.35%, recall rate of 81.20%, and precision rate of 99.99%. The proposed approach is 2.53%, 0.88%, and 2.75% higher than that of the original YOLOv3 algorithm.