基于局部敏感直方图的时空上下文跟踪

Spatio-temporal context tracking based on locality sensitive histograms

针对当前目标跟踪算法在目标区域光照剧烈变化、长时间遮挡或者平面内旋转时会发生偏移甚至跟丢这一现象,提出了基于局部敏感直方图的时空上下文跟踪算法。该算法以贝叶斯框架为基础,利用生物视觉特性,结合底层灰度特征,基于局部敏感直方图提取光照不变特征,建立目标与背景的统计相关模型来实现跟踪,使跟踪时偏移较小且不会跟丢目标。在对不同视频序列的实验表明:基于局部敏感直方图的时空上下文算法和多示例学习算法相比,在光照变化、平面内旋转或者遮挡时都表现出比较好的跟踪效果且中心误差较小,具有较强鲁棒性。

Aiming at problem that offset even lost will occur when illumination changes drastically or meets with rotation in plane and a large block, a spatio-temporal context learning tracking algorithm is proposed based on locality sensitive histogram. The algorithm is based on Bayesian framework and use biological visual characteristics based on locality sensitive histogram, extract illumination invariant features, combined with low-level gray scale features, set up statistics related model for target and background to realize tracking. Extensive experimental results show that the proposed algorithm performs good tracking effect and central error is small and robustness is strong compared with STC and MIL algorithms.