Abstract: Real-time video analysis tasks often involve running computationally intensive deep neural network(DNN) models for object tracking.In practical applications,offloading multi-stream video analysis tasks to edge devices near the cameras has become crucial.However,these edge devices often have limited computing resources,resulting in low tracking accuracy.This is primarily due to outdated detection results,accumulated tracking errors,and the inability to detect new object.To address these issues,a prediction-correction based framework is proposed.The framework comprises three core components:1)Predictive detection propagation,which rapidly updates outdated object bounding boxes using a lightweight prediction model to match the current frame.2)Frame difference corrector,which refines bounding boxes based on frame difference information.3)New object detector,which discovers newly appearing objects during the tracking process by clustering frame difference features.Experimental results demonstrate that the framework achieves accuracy improvements ranging from 19.4% to 34.7% compared to baseline methods across various traffic scenarios while maintaining real-time execution speed.

Key words: Edge device, Resource efficiency, Object detection, Object tracking