Improving Performance of CNN Based Vehicle Detection and Tracking by Median Algorithm

The performance of object tracking algorithms depends upon the results of object detectors. If object detection algorithm fails to detect the objects in the image due to occlusion, field of view, and low contrast, then the object tracking algorithm relying on detection cannot generate tracking details for those objects. Thus, we need an additional scheme that can quickly predict tracking details. We can store the previous tracking details (unique ID) and corresponding labels of each object in the history buffer. In case when object detection fails, we can use this history data to predict future labels and minimize tracking errors. In this work, we propose a median based label estimation scheme that predicts detection labels for the objects in the current frame using a median of history labels stored in the previous frames. Median has the property that it does not affect by outliers, which gives a stable value of the past and provides better estimation. Along with the predicted labels, our scheme also alleviates the problem of unstable labels showing flickering bounding boxes, and thus significantly improves the visual quality of object detectors.