基于深度学习的多目标运动轨迹预测算法

针对多目标运动轨迹预测过程中由于检测精度和实时性不足造成部分目标位置信息丢失和预测准确度不高问题,提出基于改进卡尔曼滤波的多目标轨迹运动轨迹预测(MMTP)算法。MMTP算法在目标检测阶段使用YOLOv4检测器提升目标检测的准确率和速度;在目标匹配阶段采用KM匹配算法将当前检测框的检测目标与上一时刻预测的预测框的目标进行数据关联,从而增强目标关联的准确性,避免目标遮挡、目标交错和漂移造成的目标丢失;在目标坐标预测阶段,提出改进卡尔曼滤波算法为每个运动目标预测下一帧位置坐标并画出预测框,提高非线性场景中目标坐标的预测精度,降低预测坐标的误差。使用MOT16与实际交通系统拍摄的视频序列数据集验证算法整体性能,仿真结果表明,MMTP在目标检测阶段具有较好的检测精度和速度,有效提升了算法整体的运行速度;在目标匹配阶段,MMTP算法能增强目标关联的准确性,减少目标丢失,比RMOT、POI、SORT、Deep-SORT和YVTP算法更优。

Multi-target motion trajectory prediction algorithm based on deep learning

In the process of multi-target motion trajectory prediction, due to insufficient detection accuracy and real-time performance, some target position information is lost and prediction accuracy is not high. This paper proposed a multi-target trajectory prediction algorithm(MMTP) based on improved Kalman filter. MMTP used the YOLOv4 detector in the target detection stage to improve the accuracy and speed of target detection. In the target matching stage, MMTP used the KM matching algorithm to associate the detection target of the current detection frame with the target of the prediction frame predicted at the previous moment, thereby enhancing the accuracy of target association and avoiding target loss caused by target occlusion, target interleaving and drift. In the target coordinate prediction stage, this paper proposed an improved Kalman filter algorithm to predict the position coordinates of the next frame for each moving target and drew a prediction frame to improve the prediction accuracy of target coordinates in non-linear scenes and reduce the error of predicted coordinates. Then it used the video sequence data set taken by MOT16 and the actual traffic system to verify the overall performance of the algorithm. The simulation results show that MMTP has better detection accuracy and speed in the target detection stage, which effectively improves the overall operating speed of the algorithm. In the target matching stage, MMTP can enhance the accuracy of target association and reduce target loss, which is better than RMOT, POI, SORT, Deep-SORT and YVTP.