A Motion Status Discrimination Method Based on Velocity Estimation Embedded in Multi-object Tracking

Multi-object tracking(MOT) has been a research hotspot in the field of computer vision. Currently, the most popular tracking paradigm involves first using object detection models to process the images, and then combining methods such as Kalman filtering for tracking. This paradigm has achieved good results on various datasets, but it lacks the ability to accurately determine whether an object is in motion. In certain scenarios, researchers may be interested only in moving objects and desire information specifically related to their motion, without the need for information about stationary objects. However, there has been limited research on determining the motion state of targets in the field of multi-object tracking. To address this issue, this paper proposes a velocity-based motion state discrimination method for object tracking. This method can be integrated with different multi-object tracking methods to provide more information about the tracking results. Additionally, we evaluate the effectiveness of the method on a custom dataset and test it on selected scenes of the re-annotated CDnet-2014 dataset. The experimental results demonstrate significant achievements of our proposed method. When combined with the YOLOv8+StrongSORT tracking method, the motion state discrimination accuracy (F1 score) of our method reaches 0.6975 on the custom dataset and 0.7831 on the CDnet-2014, showcasing the practicality of our method.