Real-time object detection with foreground fusion from multiple cameras using homography mapping of polygon vertices

Multi-camera and multi-plane foreground fusion approach can relieve the effects of occlusion and improve the accuracy and robustness of moving object detection. The traditional homography mapping is an image-level transformation which projects each pixel in the binary foreground image into a reference view. To avoid perspective openings or holes which are generated during the mapping from the camera view to the top view, the number of the pixels in the homography mapping is decided by the resolution of the top view, which is usually higher than that of the camera view. The slow speed has dissuaded the foreground homography mapping from real-time applications. A foreground polygon approximation method is proposed. After the foreground regions are identified in a camera view, each foreground region is approximated by a polygon and only the vertices of the polygon are projected to the reference view through homography mapping. Then the projected foreground region, which is rebuilt in the reference view, is utilized in real-time moving object detection with multiple cameras. To evaluate the performance, the proposed polygon approximation method has been compared with the contour based method and the bounding-box based method. The experimental results have shown that the proposed algorithm can produce competitive results in comparison with those using foreground bitmap mapping. Considering the differences of moving objects' size, the processing speed is about 12 and 69 times faster than the bitmap mapping method.