基于行人检测与峰值密度聚类的行人多次徘徊检测算法

Pedestrian Multiple Wandering Detection Algorithm Combining Pedestrian Detection and Peak Density Clustering

为了提升传统行人徘徊检测方法的准确性,提出了一种结合行人检测与峰值密度聚类的行人多次徘徊检测算法(Multiple Wander Detection Combining Pedestrian Detection and Peak Density ClusteringMWD_PD_DPC)。首先,在行人检测算法的特征提取网络与FPN层之间加入自适应卷积注意力机制(SKNet),提升模型在多尺度场景下行人检测精度。然后,提出了柔性非极大值抑制(DIOU-Soft-NMS)来缓解行人在密集场景下错误抑制的现象,提升行人检测算法在密集场景下的检测精度。最后,使用峰值密度聚类算法(DPC)对行人的轨迹进行分析,来判断是否发生徘徊行为。并通过AdaFace人脸识别算法对徘徊的行人进行人脸匹配,来判断行人是否在不同时间段多次发生徘徊行为。实验表明,该方法单次徘徊检测的准确率到达了94.6%。行人多次徘徊检测的准确率到达了78.7%。

In order to improve the accuracy of traditional pedestrian wandering detection methods,a pedestrian wandering detection algorithm(Multiple Wander Detection Combining Pedestrian Detection and Peak Density Clustering MWD_PD_DPC)is proposed,which combines pedestrian detection with peak density clustering.First,an adaptive convolutional attention mechanism(SKNet)is added between the feature extraction network and the FPN layer of the pedestrian detection algorithm to improve the accuracy of pedestrian detection in multi-scale scenes.Then,the improved flexible non maximum suppression(DIOU-Soft-NMS)is used to alleviate the phenomenon of false suppression of pedestrians in dense scenes and improve the detection accuracy of pedestrian detection algorithms in dense scenes.Finally,the peak density clustering algorithm(DPC)is used to analyze the tracks of pedestrians to determine whether wandering behavior occurs.And through adaface face recognition algorithm to match the faces of wandering pedestrians,we can judge whether pedestrians have wandered for many times.The experiment shows that the accuracy of single wandering detection reaches 94.6%.The accuracy rate of pedestrian multiple wandering detection reached 78.7%.