Homography based multiple camera detection of camouflaged targets

This thesis proposes an improvement on segmentation of camouflaged objects by combining information from multiple static cameras with overlapping views onto a common reference plane. Previous camouflage detection work relies on a single camera and focuses on pattern recognition, gray level co-occurrence matrix (GLCM) computation, convexity detection and operations such as connected components or morphology. These methods make real-time implementation computationally expensive. The proposed work diverges from these techniques, taking advantage of the added information from several viewpoints. Background subtraction is performed on each individual camera and then, the segmentation result is transformed onto a common plane using homography. A multiple camera view is generated by averaging the transformed views from each camera. This multiple camera view is used to achieve a better segmentation of the camouflaged object. A multi-camera surveillance system is developed in Matlab as part of this research. The surveillance system includes processes such as classification, centroid calculation, tracking using a Kalman Filter and bounding the target with a rectangle. In addition, a graphic user interface was designed to be able to choose from several cameras, viewpoints and motion detection methods. Experiments were run to test the performance of the background subtraction method versus two other reference model methods. Tests were also run to observe the challenges faced by background subtraction under camouflaged conditions. Additional experiments compare the tracking from individual cameras to the results from the Kalman Filter and Multi-Camera Average. The computation time between the proposed method and GLCM was also studied. The results indicate that a multiple camera surveillance system provides an improved segmentation of a camouflaged target and is capable of tracking targets more consistently than individual cameras. Further results indicate that a multi-camera system can perform faster than GLCM in some circumstances