基于低秩矩阵恢复交通视频背景重建性能评价

Performance evaluation of traffic video background reconstruction based on low-rank matrix recovery

为分析低秩矩阵恢复算法在交通视频背景重建中的性能,分别对基于矩阵补全、鲁棒主成分分析和低秩表示3种低秩恢复方法做交通视频背景重建实验。分别针对环境光照变化、不同车流量、阴影等场景进行测试。为客观评价算法性能,使用SBMI2015推荐的评测指标,结合算法复杂度和执行效率说明算法性能优劣。与目前常用的加权移动平均法、自适应背景学习算法和高斯混合模型法进行对比。实验结果表明,对于缓变交通场景,低秩矩阵恢复算法在处理环境光照变化、较大车流量和克服阴影等不利因素时,较传统背景重建算法有突出优势。综合指标表明,鲁棒主成分分析方法在交通场景背景视频重建中的性能优于其它方法。

For evaluating the performance of low-rank matrix recovery （LRMR） algorithms in the background reconstruction of traffic video, a series of tests were performed on matrix completion （MC） , robust principal component analysis （RPCA） and low-rank representation （LRR） to recover the traffic video background. These algorithms were compared with some traditional background recovery methods including weighted moving average, adaptive background learning and Gaussian mixture model. In the experiments, three situations were taken into account, i. e. illumination changing, different traffic flow and shadow. A group of metrics recommended by SBMI2015 was adopted as evaluation criterions, as well as the common algorithm performance metrics,such as CPU time and algorithm efficiency. Experimental results show that for the slow change traffic scene, LRMR al-gorithms outperform the traditional methods in overcoming difficult environmental conditions. In overall, RPCA has the best performance among LRMR algorithms in traffic video background reconstruction.