Small Object Tracking: A Particle Filter with
Colour Segmentation-based Detectors
Abstract: We propose an efficient method of ball localization in soccer game video integrating conventional detection and tracking. Our ball detector is based on color segmentation. We first use the likelihood ratio approach to classify the pixels by trained playfield and non-playfield color histograms and then group detected playfield pixels to connected areas. Next, a strong ball detector is created based on shape analysis on foreground blobs in these areas and used to trigger an adaptive particle filter-based ball tracker. Besides, a weak ball detector is built with outputs from the playfield detector and ball's color information and integrated into the observation likelihood of the particle filter in the tracker. In the ball tracker, motion estimation is embedded to generate a better proposal distribution and a mixture model is tailored to handle ambiguity in the ball's intensity measurement likelihood due to the cluttered background. In addition, both occlusion and template drifting are coped with explicitly. By counting the duration of continuous frames in low confidence (occlusion), the system can reboot the strong ball detector and recover from the complete tracking failure. Promising results of automatic ball detection and tracking in soccer game videos are presented to illustrate that the proposed scheme handles complete occlusion and heavy clutter effectively.
Figure 1: Playfield segmentation and foreground blob extraction
Figure 2:Diagram of "Detect-then-Detect-and-Track(DtDaT)".
Figure 3:Tracker's recovery from failure due to motion blur.
Demos
1.Video "France".
2. Video "Belgium".