Adaptive Object Tracking and Detection for Video Recognition in Mobile Edge Networks

Video recognition requires rapid and accurate object recognition across continuous frames, presenting significant challenges for computing-constrained devices such as traffic cameras. The advancements in mobile edge computing have enabled the deployment of high-accuracy neural networks on edge servers for object detection. Combined with lightweight object tracking algorithms on the local device, this approach emerges as a highly promising solution. However, determining when to perform edge detection or local tracking remains challenging. To address this, we first formulate a long-term optimization problem that accounts for the temporal correlation of frames and the dynamic nature of mobile edge networks. The achievable reward for each frame is defined as a weighted sum of recognition accuracy, handling delay, and waiting delay. Then, we propose the LTED-Ada, a deep reinforcement learning based algorithm that adaptively selects between local object tracking and edge object detection for video recognition. Finally, we conduct hardware-in-the-loop experiments to confirm LTED-Ada's superior performance across diverse frame rates and device requirements.