Sequential Monte Carlo Method for Real-time Tracking of Multiple Targets

摘要

In this project, a Monte Carlo approach to tracking was developed for tracking in cluttered environments and across multiple scales. The Monte Carlo approach was compared with an active contour approach. Specifically, we developed a novel deterministic approach for 2D projective/affine snakes, which was evaluated in conditions of high clutter and with targets of varying viewpoint and scale. A base view active contour method has been developed and tested for target tracking. The base view active contour displayed an average error 10% more accurate than the correlation tracker and 14% more accurate than the centroid tracker tested with 120 synthetic videos corrupted with both Gaussian and impulse noise. Over 46 real video sequences base view active contours successfully tracked the target in an average of 80% of the frames as compared to 73% of the frames for the centroid tracker and 83% for the correlation tracker. When the real video sequences containing target occlusion were removed from consideration, the base view active contour successfully tracked in an average 87% of the frames whereas the correlation tracker's performance dropped to only 75% of the frames. Overall, base view active contours outperform the competing methods in the synthetic and real video experiments.