综合多种预测方案实现遮挡情况下的目标跟踪

由于采用了多种运动预测方案,本文提出的目标跟踪方法能选择最佳的观测结果,实现对非标定固定焦距的静止摄像机的单目图像序列中的目标跟踪.静态背景参考图像由混合模型法估计,在最简单的环境下,跟踪算法则采用匀加速运动模型对目标完成跟踪.本文主要贡献是采用了三个预测器和最小方差相关时选择目标最可能的位置.三个预测器分别是:-跟踪方案、卡尔曼滤波和区域分割匹配方案.本跟踪方案通过具有不同遮挡情况的序列图像得到了验证.

Occlusion Tolerent Tracking Using Hybrid Prediction Schemes

A method of combining multiple moving objects prediction schemes is presented that allows a tracking framework to select and identify the best observation evidence in occlusion scenarios. The underlying framework tracks any objects in monocular image sequences taken from stationary uncalibrated cameras with fixed focal length. A mixture model method is deployed to estimate the static background reference image. The tracking algorithm simply uses a constant acceleration motion model to track objects in the simplest scenarios. However, the main contribution is the use of three simultaneous predictors with a least square correlation stage to select the most likely object position. The three prediction schemes are an α-β tracking scheme, a Kalman filtering method, and a region segmentation and matching method. The tracker is evaluated against different image sequences each offering different occlusion problems.