基于时空域暗示的运动对象检测和分割*

通过对视频帧间的时域信息和帧内的空域信息的充分开发，提出一种任意视频对象检测与分割算法。初始的运动区域评估利用时间加权的时域帧窗，采用基于点的分割；而近似同质颜色亮度纹理区域利用区域之间的差异和区域内的相似，采用改进的分水岭分割和基于区域特征相似度的合并。时域和空域分割结果的合并基于多数原则。最后，分割结果的完善和修正基于时域的持续性和空域上的一致性标准。通过测试，提出的分割算法获得可靠的对象边界，而且通过调整少量参数，可以适应于室内和室外场景以及高速和低速运动物体。

Detection and segmentation of moving objects using temporal and spatial cues

This paper proposed a novel scheme for automatic video objects detection and segmentation by fully exploiting both inter-frame temporal information and intra-frame spatial information from video stream. In the proposed method, detected the initial motion regions by using a time-weighted temporal window, and segmented many approximately homogenous color-texture regions by improved watershed and spatial similarity criterion, built further merging results of temporal and spatial segment available on majority rule, refining them was based on spatial coherency and temporal continuity eventually. The outputs show that the proposed scheme is robust and gets fine boundary of objects. Furthermore, it is adaptable to high-speed and low-speed objects in indoor and outdoor scene by training and adjusting few parameters.