Panoramic Multi-Target Tracking Model Incorporating Human Features and Motion Model
Jinfeng Shen, Hongbo Yang, Wenfeng Song · 2024
With the continuous innovation in deep learning methods, many novel algorithms have been proposed in the fields of object detection and tracking, achieving impressive results. Extensive experimental validation has revealed that existing tracking algorithms face issues with target loss at the boundaries in panoramic scenes. To address this problem, a panoramic multi-object tracking model that integrates human features and motion models was proposed. This model combines the Reid method for target matching based on human features and employs an improved Kalman filter for predicting target trajectories based on motion models. The algorithm also includes a camera motion compensation module in the tracking process to mitigate the impact of potential camera shake on tracking results. Through experimental validation, the new algorithm demonstrated its ability to accurately track pedestrian targets at the boundaries. Performance evaluation on the test set showed tracking metrics of IDF1 =78.1, higher than other existing algorithms, and IDSW=820, lower than existing algorithms, indicating superior performance.