Abstract
An automatic surveillance system should be able to detect events in which users are interested in. Due to the unique characteristics of highway environment, event detection becomes much easier. In this thesis, we propose an approach to detect events in highway surveillance environment. First, for each pixel, we adopt GMM to construct background model and detect foreground. Then, we adopt two-stage voting approach for blob matching and modify the voting conditions based on video properties. Besides, we also conduct occlusion detection and correct those occluded blobs caused by segmentation error. From accumulative tracking data, blob center histogram is constructed to detect trajectory curves. We divide the histogram into sub-images, use Hough transform to find the initial position, apply second-order least-square for curve approximation, and reconstruct lanes. With lane information, we modify the occlusion correction method and detect lane-crossing event. The experiment results show that the modified occlusion correction method can retrieve more precious minimum bounding box position for the occluded blobs and get better tracking results.