Unsupervised classification of erroneous video object trajectories

The paper proposes a method to detect failures in object tracking. Detection is done with the help of two types of errors, namely jump and stop errors. Jump errors occur when an abrupt change in object’s motion is observed, whereas stop errors occur when a moving object remains stationary for longer duration at any point. In our framework, moving objects are first tracked using well-known trackers and their trajectories are obtained. Discrepancies between trajectories are measured. We have shown that the proposed method can be reliable for detection of tracking failures. This can help to find error-free trajectories that are essential in various computer vision tasks. We have also shown that the tracking performance can be further improved while processing the output trajectories without much knowledge about the underlying tracking algorithms. The effect of tracking failure is investigated to identify erroneous trajectories. It has been observed that when a tracker fails, velocity profile of the moving object usually changes significantly. Based on this hypothesis, erroneous trajectories are detected and a set of error-free trajectories are marked and grouped. Two recently proposed tracking algorithms, namely real-time compressive tracker (CT) and real-time L1-tracker (L1APG), have been used to track the objects. We have tested our framework on five publicly available datasets containing more than 300 trajectories. Our experiments reveal that average classification rate of erroneous trajectories can be as high as 80.4% when objects are tracked using L1APG tracker. Accuracy can be as high as 81.2% when applied on trajectories obtained using CT tracker. Average accuracy of tracking increases significantly (19.2% with respect to L1APG tracker and 24.8% with respect to CT tracker) when the decision is taken using a fused framework.