Flexible tracklet association for complex scenarios using a Markov Logic Network

Multi-object tracking, despite significant advances in recent years, is still an area where improvements can be made. The additional constraint of the system running in real-time further limits the complexity of the algorithm. Consequently, real-time multi-object tracking algorithms are in general not sufficiently robust to produce error-free outputs. However for offline applications, the track quality can benefit from a post-processing step. We consider such an approach in the form of tracklet association, where reliable track fragments are joined to form longer, coherent tracks. This is implemented with a Markov Logic Network.