Proceedings of the 29th IEEE Conference on Computer Vision and Pattern Recognition Workshops, 2016.
29th IEEE Conference on Computer Vision and Pattern Recognition Workshops, Las Vegas, Nevada, June 26 - July 1, 2016
Multiple object tracking in Wide Area Motion Imagery (WAMI) data is usually based on initial detections coming from background subtraction or frame differencing. However, these methods are prone to produce split and merged detections. Appearance based vehicle detection can be an alternative but is not well-suited for WAMI data since classifier models are of weak discriminative power for vehicles in top view at low resolution. We introduce a moving vehicle detection algorithm that combines 2-frame differencing with a vehicle appearance model to improve object detection. Our main contributions are (1) integration of robust vehicle detection with split/merge handling and (2) estimation of assignment likelihoods between object hypotheses in consecutive frames using an appearance based similarity measure. Without using any prior knowledge, we achieve state-of-the-art detection rates and produce tracklets that considerably simplify the data association problem for multiple object tracking.