基于TLD改进的自动人体检测与实时跟踪算法  

An automatic human detection and real-time tracking algorithm based on TLD

作　　者：毛家顺 张汝波 杨大伟 

机构地区：[1]大连民族大学机电工程学院,辽宁大连116600

出　　处：《微型机与应用》2015年第22期47-49,共3页Microcomputer & Its Applications

基　　金：大连金州新区科技计划项目(KJCX-ZTPY-2014-0005);大连民族大学中央高校自主基金(DC201502010304)

摘　　要：首先提出一种运动人体检测算法,通过图像序列识别出运动人体作为跟踪目标,然后在TLD算法中引入目标轨迹预测,利用该信息来辅助空间搜索。运动人体检测算法首先采用背景减除和边缘检测算法获得完整目标轮廓,然后使用HU仿射不变矩检测出运动人体,该特征能适应目标旋转、尺度、仿射等变化场合。为提高跟踪实时性,在TLD框架中引入卡尔曼轨迹预测,并先在预测位置邻域搜索。实验结果表明,该运动人体检测算法能够在静态和动态背景下准确地检测出运动人体;改进后的TLD算法与原始算法相比,在准确率不降低情况下,降低了计算复杂度。In this paper, firstly, we propose an automatic human detection algorithm, which is used to detect the moving human body from image sequence as tracking target, and then predict the target motion state and use this information to assist target search. The Caiman filter is introduced into the TLD frame. The automatic human detection algorithm uses background subtraction and edge detection to obtain complete contour of the moving object from image sequence, then discriminates this object according to the contour region features, which can adapt to the change of the object＇s rotation, scale, affine, etc. In order to improve the realtime performance of tracking algorithm, the object trajectory prediction is introduced into the TLD tracking algorithm. To simplify search space, the detector firstly searches the object within predicted position. Experimental results show that the proposed human detection algorithm can accurately detect human object, and compared with the original TLD algorithm, the improved TLD algorithm can reduce the computational complexity in the case of constant accuracy and track in real-time under static and dynamic background.

关 键 词：背景减除 人体检测 边缘检测 卡尔曼滤波器 TLD 

分 类 号：TP391.41[自动化与计算机技术—计算机应用技术]