Detection and ReID are Not All you Need: Accelerate Multi-Object Tracking with Correlation Filter

Tracking by detection (TBD) is a typical paradigm in multi-object tracking (MOT) tasks. This paradigm relies on accurate object detection models and object re-identification (ReID) models to detect objects in videos and associate them with tracklets. However, current MOT algorithms are constrained by the parameter size of these models, leading to high computational costs. Additionally, using real-time object detection models with average accuracy can result in issues like ID switching and tracklet interruptions. Single-object tracking (SOT) methods based on correlation filters, which utilize frequency domain operations and kernel functions to optimize the computation process, are suitable for balancing the efficiency and performance of the TBD paradigm. This paper designs two efficient, plug-and-play modules based on correlation filters. First, a correlation filter interpolation tracking module is introduced, with a well-designed strategy to intermittently use correlation filters to track each existing tracklet, replacing the object detection and ReID models in the TBD paradigm. This approach retains the original paradigm's ability to detect new objects while leveraging the speed and robustness of single-object tracking. Additionally, the paper establishes a correlation filter-based object re-identification module, replacing the slower ReID model by leveraging the correlation responses. Experiments demonstrate that the proposed method significantly improves the real-time performance of tracking algorithms. Furthermore, it enhances the accuracy of tracking algorithms when using real-time detectors with average performance.