A multi-object tracking method based on global-local feature joint modeling

Abstract

This invention proposes a multi-object tracking method based on joint global-local feature modeling. This method generates a multi-scale image pyramid for the current frame of a large-scene, high-resolution video, obtaining multiple multi-scale image representations at different resolutions. A sliding window approach is used for object detection, and a non-maximum suppression algorithm is employed to fuse the detection results at each scale, constructing a joint query group containing global and local object queries. This joint query group is input into the decoder and associated with the encoded image features through a cross-attention mechanism, outputting a joint global-local feature representation. Combining the target's occlusion state prediction results, an occlusion state-aware matching strategy is used to optimally match the currently detected target with the trajectory set, dynamically updating or discarding trajectories. This invention effectively improves the tracking accuracy and continuity of multiple objects in large-scene, high-resolution videos under dense occlusion environments, achieving collaborative optimization of global and local features.

Images