Joint Frame Drop and Object Detection Task Offloading for Mobile Devices via RL With Lyapunov Optimization

Vaughn Sohn, Suhwan Kim, Hyang-Won Lee

Abstract: Object detection has become an increasingly important application for mobile devices. However, state-of-the-art object detection relies heavily on deep neural network, which is often burdensome to compute on mobile devices. To this end, we develop a layering framework for joint video frame drop and object detection task offloading. In the lower layer, by invoking Lyapunov optimization, we devise an algorithm for partitioning and offloading the computation tasks of deep neural networks. This algorithm also specifies the flow control for admitting the application traffic into the network. In the upper layer, we use the flow control as a form of guidance in the action space in order to develop a reinforcement learning (RL) algorithm that selectively drops video frames with object detection performance in consideration. By the nature of design, this Lyapunov-guided RL guarantees the network stability. We show through simulations that our Lyapunov-guided RL drops video frames with reasonable object detection performance and reduced latency while keeping the network stable. We also implemented our algorithm on the remote-controlled (RC) car equipped with microprocessor and GPU, and demonstrate the applicability of our algorithm to real-time object detection tasks from the video stream generated as the RC car moves.

External IDs:dblp:journals/tmc/SohnKL25