SIPA培训教程文件.pptVIP

Chang Liu, George Chen, Yingdong Ma Shenzhen Institutes of Advanced Technology Yajun Ding Hunan Normal University Jinwu Wu Shenzhen Transportation Depot Group Accurate Vehicle Tracking by Shadow Removal in the Transformed Space System Overview Problems Solutions Experimental Results Conclusions Future Work Outline 2 System overview -- Flowchart 3 Frame in raw space Final result Result in transformed space Frame in transformed space System overview -- Example (1) Frame in raw space (2) Background in transformed space (3) Foreground in transformed space (4) Final result in raw space 4 Incorrect tracking results, two vehicles tracked as “041” because of perspective effect Problem1 -- Perspective effect (5) (6) 5 Two vehicles tracked as a target !!! Solution1 -- Perspective transform (7) Characteristic points in raw frame (8) Corresponding points in expected frame 6 (1) (2) Solve H matrix The two vehicles tracked as “048” and “053” respectively Result1 (9) (10) 7 Two vehicles tracked as two targets Problem2 -- Shadow (11) Frame in raw space (12) Background in transformed space (13) Foreground in transformed space (14) Vehicles are connected 8 Solution2 -- Color difference edge difference (15) Background ROI (16) Foreground ROI (17) Color difference (18) Edge difference 9 + Or Result2 (19) Moving vehicle mask (20) Vehicles are separated 10 Single vehicle tracked twice as “003” and “005” separately Problem3 -- Bounding rectangle histogram (21) (22) 11 Masked histogram - Use foreground mask to construct the histogram Solution3 -- Masked histogram (23) (24) 12 The vehicle tracked only once Result3 (25) (26) 13 Correct tracking result - Vehicle tracked as “003” invariably because of the masked histogram Comparison of different methods Video 1 Video 2 Video 3 Frames 2699 1994 9400 Actual Number of Vehicles 232 56 526 MS # detected 161 39 331 Accuracy 69.40% 69.64% 62.9% MSFG # detected 196 48 429 Accuracy 84.48% 85.71%