基于分水岭分割和尺度不变特征点的多目标全自主跟踪算法

该文针对多目标的鲁棒跟踪问题，设计了一种基于图像分水岭分割和尺度不变特征变换(SIFT)特征点的多目标全自主跟踪算法。为规避图像平坦区域，提出在原图上叠加规则坡度图的思想，并在浮点域进行一定尺度高斯模糊处理，将区域极小值点作为种子点完成分水岭分割，并将极值点作为目标特征点，通过前后帧分水岭映射生成特征点短时轨迹，自动检测运动目标。之后依据目标所处状态(是否发生遮挡)和分水岭分割图建立、更新目标SIFT特征池，结合分水岭映射、SIFT特征池匹配完成对目标的鲁棒跟踪。实验结果表明，该算法能有效完成视频中多目标的持续跟踪，并对目标遮挡有较好的鲁棒性。

Full-automatic Tracking Algorithm for Multi-object Based on Watershed Segmentation and Scale-invariant Feature Points

For the issue of multi-object robust tracking, a type of watershed segmentation and Scale-Invariant Feature Transform (SIFT) feature points based full-automatic tracking algorithm is presented. To avoid flat area while do watershed segmentation on the image, a regular gradient image is added to the source image. After the Gaussian blurred process is done on the added image in float field, field minimal points are selected as object feature points as well as seed points to do watershed segmentation. Moving object is detected through short time points trajectories derived from watershed region mapping relationship between current and backward image. SIFT feature pool is built and updated based on object occlusion occurred or not and watershed segmentation. With the help of watershed region mapping and feature matching with the SIFT feature pool, object is robustly tracked. Actual tests show that the algorithm can track multi-object well and with a better performance of mutual occlusion robustness.