Résumé
The Generalized Labeled Multi-Bernoulli (GLMB) filter attains remarkable results in Multi-Object Tracking (MOT). Nevertheless, the GLMB filter relies on strong assumptions such as prior knowledge of targets' initial state. Pragmatic scenarios such as satellite video object tracking challenge these assumptions as objects appear at random locations and object detectors output numerous false positives. We present an enhanced version of the GLMB filter that learns from previous trajectories to estimate accurate hypotheses initializations. We keep track of previous target states and use this information to sample the initial velocities of newborn targets. This addition significantly improves the performance of the GLMB in videos with low Frames Per Second (FPS), where the target's initial states are paramount for object tracking. We test this enhanced GLMB filter versus comparable trackers and previous solutions for the GLMB filter and show that our filter obtains better performance. Code is available at https://github.com/Ayana-Inria/GLMB-adaptivebirth-satellite-videos