Aiming at the shortcomings of traditional cloud-based visual monitoring system in real-time and privacy protection, this paper proposes a key technical framework of real-time visual monitoring system that integrates edge computing and deep learning. The system adopts "end-edge-cloud" three-level collaborative architecture, and optimizes the task allocation between edge and cloud based on reinforcement learning (RL) through adaptive model segmentation and dynamic unloading algorithm, which significantly reduces the end-to-end delay. A hardware-aware hybrid pruning method is designed, which combines structured and unstructured pruning strategies to orient the compression model according to the characteristics of edge devices, so as to improve the reasoning speed and reduce energy consumption while maintaining high accuracy. The key frame priority scheduling mechanism is introduced, and the abnormal events can still be responded quickly under the high load of the system through ROI (Region of Interest) pre-extraction and asynchronous calculation. The experimental results show that compared with the traditional scheme, the proposed method can achieve better balance of delay, accuracy and energy consumption under different network conditions, and has good adaptability and practicability, which provides a feasible technical path for real-time visual application scenarios such as industrial quality inspection and intelligent security.
Li et al. (Sun,) studied this question.