Robust Segmentation of Freight Containers in Train Monitoring Videos Qing-Jie Kong*, Avinash Kumar**, Narendra Ahuja**,Yuncai Liu* **Department of Electrical and Computer Engineering University of Illinois at Urbana-Champaign *Institute of Image Processing and Pattern Recognition Shanghai Jiao Tong University, Shanghai , China WACV 2009
Input Input : Video of an Intermodal Freight Train captured from a fixed camera with background visible before the train arrives.
Camera Inter-modal Train Viewing Volume Video Capture of an Intermodal Freight Train
Output Output : Video with background removed and thus foreground consisting of only the intermodal train. Main Application : Fast and Automatic calculation of gap lengths between consecutive containers
Example: Input and Output Video
Major Difficulties Varied Outdoor Imaging Conditions
Major Difficulties Different Types of Containers
Four Stage Coarse-to-Fine Framework Stage 1: Detecting Train Region Stage 2: Removing Background Gap Stage 3: Detecting Single Stack Stage 4: Refining Segmentation Result
Stage 1: Detecting Train Region Partition of the region Pixel signal in temporal domain Power spectrum of the signal
Stage 1: Detecting Train Region A frame in a video Frequency image of the video Histogram of the frequency image
Stage 1: Train Region Detection Thresholded result By the morphological operations Final result of the first stage
Stage 2: Removing Background Gap Background Model Background Removal Background Update
Background Model A background image A sub-region of the background image Histogram of the sub-region
Background Removal A frame in a video Result of the recognition Segmentation result
Background Update The background between two containers spans the complete background every some frames Splice the detected middle backgrounds to rebuild a new background image The updating calculation happens as soon as the middle background region completes the scan
Stage 3: Detecting Single Stack A frame in a video Result after the first two stages blobSegmentation result
Stage 4: Refinement of Segmentation Results Result before refinement Result after refinement Background image A window of background
Combination of Color Information Do all the processing in Stage 2 to the RGB channels respectively Combine the results of the three channels by the AND operation
Experiments Video data: 150 videos Include 1222 containers and a wide range of background conditions: – clear blue sky – bright sunlight – static heavy clouds in the day and evening – moving heavy clouds in the day and evening – rainy day (water on lens)
Experiments Success ratio of Stage 1: 96% Success ratio of the last three stages: Background Conditions Total Numbers of the Containers SR (Grey)SR (RGB) 1. Day/No cloud/Blue sky %96.2% 2. Day/Bright sunlight6188.9%91.8% 3. Day/Heavy clouds %100.0% 4. Day/Moving cloud %100.0% 5. Day/General situation %100.0% 6. Evening/heavy clouds % 7. Evening/Moving clouds % 8. Rainy day/Water on lens % Total %99.3%
Experiments
Operation speed – computer : Intel(R) Core(TM)2 Due CPU 2.53-GHz processor and 3.2-GB MHz RAM. – average processing speed: 4 frames per second (fps)
Conclusion The proposed method – combines the information in frequency and spatial domain – is robust to varieties of background conditions – can employ videos from un-calibrated cameras – is being integrated into a real time vision system for intelligent train monitoring
Thanks to BNSF