基于类锚虚拟线圈的多流向车流量检测算法

针对城市道路车流量检测中车辆误分类问题，提出一种基于类锚虚拟线圈的多流向车流量检测算法。首先，采集车辆图像样本并随机裁剪以构建小客车、公交车和摩托车的均衡数据集，通过 DBSCAN(Density-Based Spatial Clustering of Applications with Noise)算法聚类获得 3 类车型的高度、宽度尺寸，以此校正场景车辆识别线圈尺寸，布设物体检测线圈与组合车辆识别线圈；其次，基于均衡数据集训练ResNet18卷积神经网络完成车辆类型判断；最后，采用改进的核相关滤波器追踪算法追踪车辆轨迹，通过计数线完成多流向车流量检测。验证分析表明：对单向车流，高峰、平峰正检率均值提升了5.09%、4.57%，误检率均值降低了5.31%、2.35%；多向车流中，直行车流的高峰、平峰正计率提升了 5.01%、5.99%，左转车流的高峰、平峰正计率提升了 4.29%、 4.56%。

A Multi-direction Traffic Flow Statistics Algorithm Based on Anchor-like Visual Loops

To reduce the problem of vehicle misclassification in urban road traffic flow detection, a multi- direction traffic flow detection algorithm based on anchor-like virtual loops is proposed. Firstly, the vehicle image samples are collected and randomly divided to construct the balanced data set of passenger cars, buses, and motorcycles. The height and width of three types of vehicles are obtained by clustering with the DBSCAN algorithm, so as to correct the size of scene vehicle recognition loops, set object detection loops, and combined vehicle recognition loops. Secondly, the ResNet18 convolutional neural network is trained based on balanced data set to distinguish vehicles. Finally, the improved kernel correlation filter tracking algorithm is used to track the vehicle trajectory, and the multi- direction traffic flow detection is completed through the counting line. The experiment analysis shows that for one-way traffic flow, the average positive detection rate at peak and non-peak hours increases by 5.09% and 4.57%, and the average false detection rate decreases by 5.31% and 2.35%. In the multi-directional traffic flow, the positive counting rate for straight traffic flow at peak and non-peak hours is increased by 5.01% and 5.99%, while that for left-turn traffic flow increased by 4.29% and 4.56%.