Spatiotemporal Activity Based Moving Target Matching through Long Occlusions and Merge-Split

In this paper, we consider the problem of matching moving targets between two video sequences recorded by different stationary uncalibrated cameras simultaneously with overlapping views, without imposing any constrains on, and any prior information of, orientations, epipolar geometry, camera locations or resolutions are not required. However, the traditional color-based matching method and feature-based matching method are doomed to fail especially when Occlusions and Merge-Split exist. Ermis et al. [10] proposed an activity feature which has geometry independence to perform pixel-level matching between the cameras. We first introduce a spatiotemporal feature of target which is consists of trajectory of target and activity feature. Second, we propose a target association method in single camera with long occlusions and Merge-split. Unlike the temporal activity-based matching method applies exhaustive pixel-wise matching which is time-consuming, we just require one match computation for the duration of the video. Visualized comparison with the scale-invariant feature transform (SIFT) and the temporal activity-based matching method. The experiments show that our method can get accurate match even though split-merge exists.