Improvement of MOSSE Object Tracking Using YOLO and Adaptive Multi-Filter

Abstract

This research aims to improve the object tracking algorithm, Minimum Output Sum of Squared Error (MOSSE). MOSSE is one of the correlation filter-based object tracking algorithms. Its processing speed is fast and easy to implement. However, it is easy to lose the target during tracking because of the deformation, rotation, and occlusion of the object. Two methods were proposed to improve MOSSE. They are re-tracking mode and adaptive multi-filter. Re-tracking mode utilizes YOLO to search the candidate objects on the image once the target is lost. The correlation would be manipulated between the filter and the possible object to determine the position of the lost target according to the peak to sidelobe ratio. In addition, adaptive multi-filter method cropped different images of the target appearance in the video to get the multiple templates. More than one filter was generated based on these templates to improve the tracking performance of MOSSE. These filters were created based on the change of the target appearance such as deformation, illumination variation, and rotation, which would cause the target lost because one filter could not satisfy all the target appearance in the video. The tracking performance was tested by 27 videos in the Object Tracking Benchmark 50 (OTB-50) database. The experimental results showed that the Spatial Robustness Evaluation (SRE) was improved from 0.215 to 0.291, and the Temporal Robustness Evaluation (TRE) was from 250 to 0.346 with the re-tracking mode and adaptive multi-filter.