Moving object detection receives much research interest in contemporary computer vision studies.It is also widely acknowledged that many problems, including illumination changes, partial or full occlusions, rigid or non-rigid shape transformation, are still challenging and hinder the detection performance from being further improved.In this paper, a novel algorithm of moving object detection is introduced.The incremental learning technique is employed in this algorithm;whose main purpose is to incorporate high spatial correlation within individual frames as well as high temporary correlation between consecutive frames for automatic updating of the detection model.Also, the model learning is realized via spectral clustering.A databased composed of over 1000 video frames is utilized for experimental evaluation.A series of statistical analysis, including ANOVA and post-hoc multiple comparison tests, are implemented to evaluate the new algorithm and other compared methods.It turns out that the novel algorithm can outperform others in terms of detection accuracy and robustness from the statistical perspective.%运动目标物体检测是计算机视觉领域的热门研究方向之一.该方向的一些复杂问题,例如:环境光照变化、目标物体部分/全遮挡、目标物体刚性/非刚性形变等,仍极具挑战性,并制约检测算法效果的进一步提高.为此,提出了一种新颖的运动目标物体检测算法.该算法采用了增量学习技术,融合了视频相邻帧在空间和时间上的高相关性,在每个测试帧上都利用其相邻帧的训练数据进行模型的自学习与更新,从而保证了模型在不同环境或复杂背景下能自动调整.为了实现模型学习,还提出并采用了一种新颖的谱聚类技术.该算法通过一个由1 000多帧的视频数据库验证,采用统计学中的方差分析和多重对比等实验手段,综合分析了该算法与其他同类经典算法的效果.通过大量统计分析,结果表明,该新颖检测算法比传统算法在运动目标物体检测的准确性和鲁棒性上都有明显提高.
展开▼
