دانلود کتاب استفاده از تشخیص‌های پنهان و اطلاعات هدف در ردیابی چند نفره

Multi-person tracking is a fundamental task in computer vision with various applications such as surveillance, autonomous driving, and sports analysis. The goal is to localize and identify all persons in each frame of a video sequence. This allows to track persons in safety-critical areas, predict the movement of pedestrians in road traffic, or calculate running statistics at soccer games. The majority of methods follows the tracking-by-detection paradigm dividing the tracking problem into the two subtasks detection and association. For the generated detections, motion and appearance cues are typically extracted to solve the association task of joining detections from the same targets to tracks. This thesis shows that existing tracking approaches from the literature exploit this available information in an insufficient way. Consequently, a novel tracking framework is introduced that improves both the utilization of available detections as well as the fusion mechanism of motion and appearance information in the association. Most tracking errors occur in crowds, where missed detections due to occlusion complicate the association task. To improve the performance in such situations, an adapted non-maximum suppression is proposed, which allows to include detections under severe occlusion in the association process that were discarded by previous tracking approaches. Two different techniques are suggested to leverage the additional set of heavily-occluded detections. The first one integrates these in a second association stage, where they are matched to the remaining unassigned tracks from the first stage. The second approach utilizes the available track information to identify track clusters with missing detections and incorporates the heavily-occluded detections in these areas locally. Both techniques not only enhance detection recall under strong occlusion but also simplify the association task by reducing the number of missing detections and thus eliminating ambiguities in the assignment of detections to tracks. Next to the usage of detections, the way of leveraging the available information on appearance and motion of targets plays a key role for the association accuracy. In this thesis, existing fusion mechanisms for motion and appearance information are evaluated within a common base framework, allowing a thorough and fair comparison for the first time, and weaknesses of the prevailing approaches are elaborated. Building on this, novel combined distance measures for a better utilization of motion and appearance information in the association are introduced that significantly outperform previous variants. To prevent the start of ghost tracks from duplicate detections in crowded areas, an occlusion-aware initialization strategy is suggested. It derives knowledge about the neighborhood of unassigned detections from the available track information to identify and discard duplicates. Moreover, a lightweight model for compensating potential camera motion is presented, which is of great importance for applications with non-static cameras. The proposed modules are combined into a novel framework that surpasses the state of the art in established multi-person tracking benchmarks. This achievement is mainly due to a better use of available information in the tracking process, since the same models are adopted for detection as well as extraction of appearance and motion information as in the competing trackers. Additionally, several optimizations are made to accelerate the computationally complex multi-person tracking framework, including the application of an efficient model for extracting appearance information, the use of a high performance library for neural network inference, and parallelization. Without a significant loss of performance, the resulting system runs in real time while being capable of tracking hundreds of targets simultaneously. This is despite the fact that it includes all important components like appearance model or camera motion compensation, which are not used by many methods from the literature in order to achieve a low runtime.