基于谱聚类和增量学习的运动目标物体检测算法研究

发布时间：2018-11-12 10:33

【摘要】：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。
[Abstract]:Moving object detection is one of the hot research directions in the field of computer vision. Some complex problems in this direction, such as environmental illumination variation, partial / total occlusion of target object, rigid / non-rigid deformation of target object, etc., are still very challenging, and restrict the further improvement of detection algorithm. Therefore, a novel object detection algorithm is proposed. The algorithm adopts incremental learning technology, which combines the high spatial and temporal correlation of adjacent frames, and uses the training data of adjacent frames for model self-learning and updating in each test frame. This ensures that the model can be adjusted automatically in different environments or complex backgrounds. In order to realize model learning, a novel spectral clustering technique is proposed and adopted. The algorithm is verified by a video database of more than 1,000 frames, and the effect of this algorithm and other similar classical algorithms is analyzed synthetically by means of variance analysis and multiple comparison in statistics. Through a lot of statistical analysis, the results show that the accuracy and robustness of the new detection algorithm are obviously improved compared with the traditional algorithm.
【作者单位】：南昌大学信息工程学院;江西农业大学软件学院;华中光电技术研究所;西安通信学院;
【基金】：国家自然科学基金(61403182、61363046)资助
【分类号】：TP391.41

【相似文献】