Model-Free Multiple Object Tracking with Shared Proposals

Most previous methods for tracking of multiple objects follow the conventional “tracking by detection” scheme and focus on improving the performance of category-specific object detectors as well as the between-frame tracklet association. These methods are therefore heavily sensitive to the performance of the object detectors, leading to limited application scenarios. In this work, we overcome this issue by a novel model-free framework that incorporates generic categoryindependent object proposals without the need to pretrain any object detectors. In each frame, our method generates a small number of target object proposals that are shared by multiple objects regardless of their category. This significantly improves the search efficiency in comparison to the traditional dense sampling approach. To further increase the discriminative power of our tracker among targets, we treat all other object proposals as the negative samples, i.e. as “distractors”, and update them in an online fashion. For a comprehensive evaluation, we test on the PETS benchmark datasets as well as a new MOOT benchmark dataset that contains more challenging videos. Results show that our method achieves superior performance in terms of both computational speed and tracking accuracy metrics.