Hyper DeepSORT: Elevating Precision in Multi-Object Tracking through HyperNMS and Adaptive Kalman Filtering Innovations.

Multiple Object Tracking (MOT) aims to employ computer vision techniques for real-time tracking and recognition of multiple objects within video sequences. It encompasses the tasks of detection, tracking, and Re-identification (ReID) of objects to achieve continuous tracking of targets over both temporal and spatial domains. MOT makes up a significant challenge within the domain of computer vision. This paper proposes Hyper DeepSORT, an advanced MOT model integrating three significant innovations: HyperNMS, Hyper Kalman Filter, and MTRNet. HyperNMS, a novel Non-Maximum Suppression (NMS) technique, leverages parallel matrix operations to perform NMS in a single iteration, enhancing object recognition accuracy and system efficiency. The Hyper Kalman Filter, an adaptive variant of the traditional Kalman filter, dynamically adjusts noise covariance based on detection confidence, improving the tracker's adaptability and robustness. Additionally, MTRNet incorporates ReID technology to refine feature representation within the DeepSORT framework, encompassing attributes like colour, texture, shape, and motion parameters, bolstering tracking performance. Experimental evaluations on multiple MOT datasets show Hyper DeepSORT outperforms existing models. Specifically, it shows average improvements of 12.75%, 5.37%, 7.20%, 9.94%, 4.90%, and 12.25% over current mainstream models in mAP, MOTA, IDF1, IDSW, FP, and FN metrics, respectively. These results underscore Hyper DeepSORT's superior accuracy and efficiency in complex tracking scenarios. [ABSTRACT FROM AUTHOR]