基于卡尔曼滤波-LSTM模型的车速估计方法

Vehicle Speed Estimation Method Based on Kalman Filter-LSTM Model

为提高基于视频的道路车速统计效率,提出一种基于卡尔曼滤波-LSTM模型的车速估计方法。首先,采用YOLOv4构建车辆目标检测算法,将主干网络改进为MobileNetV2以提高目标特征的提取速度;然后,针对卡尔曼滤波算法在复杂行车场景下的车速统计误差较大的问题,结合LSTM模型对目标跟踪结果进行检测融合;最后,根据感兴趣区域内多目标跟踪的距离与时间来估计机动车速度,并对该方法进行误差分析。在长沙市枫林三路信号交叉口的监控视频进行了应用,结果表明:提出的速度估计方法有效改善了跟踪对象标签频繁切换的现象,车速检测准确率达到94.5%。

In order to improve the video-based statistic efficiency of vehicle speed at urbanroads,a speed estimation method based on Kalman filter and LSTM model is proposed.Firstly,the vehicle target detection algorithm is constructed by YOLOv4,and the backbone network is improved to MobileNetV2 for feature extraction speed.Then,to address the problem of large statistical error caused by kalman filter tracking algorithm under complex road scenarios,the LSTM model is used to fuse the tracking result;finally,the vehicle speed is estimated according to the distance and time of multi-target tracking results in the region of interest,and the error of this method is analyzed.The proposed method is applied to the surveillance video of Fenglin 3 Road in Changsha.The results show that the proposed method effectively improves the frequent ID-switch problem during target tracking,and the accuracy of the speed statistic reaches 94.5%.