基于AdaBoost和RVM的实时多目标跟踪

摘要

A multiple target tracking algorithm based on AdaBoost detection and RVM regression——AdaBoost-RVM algorithm is proposed.The proposed algorithm can handle the multi-target tracking problem with variable number of targets.An AdaBoost detector for recognizing the targets is trained.All targets are detected which existed in the scene for several frames interval, and the HSV spatial histogram and the HOG feature are matched with those targets which have been tracked.To determine whether has target disappeared, reappeared, or new target appeared.For new target, a sample set is constructed, using the RVM training, and then the RVs is obtained for predicting.Additionally, by reducing the dimensions of the sample features, thus the raining time and the prediction time are reduced.The Kalman filtering is using to revise the prediction value, which improved the tracking accuracy.This algorithm is applied to pedestrian tracking and vehicles tracking.Experimental results show that the proposed algorithm can effectively handle the problems such as variable number of targets, target occlusion and so on, and achieve real-time multi-target tracking.%提出了基于AdaBoost检测和相关向量机回归预测的多目标跟踪算法--AdaBoost-RVM算法,该算法可以处理目标数可变的多目标跟踪问题.训练一个AdaBoost检测器,用于识别目标,间隔一定帧数检测场景中出现的所有目标,并与已跟踪目标的HSV空间直方图和HOG特征进行匹配,以确定目标为消失,重现或新目标出现,对新目标构建样本集,采用RVM训练得到用于预测的相关向量.另外,缩小样本特征维度从而缩短了训练和预测时间,采用卡曼滤波对预测值进行修正,提高了跟踪精度和稳定性.将其应用于行人和汽车的跟踪,实验结果表明,该算法能有效地处理跟踪过程中的目标数变化,目标遮挡等问题,实现多目标的实时跟踪.