Edge Computing for Real-Time Video Analysis: Unloading Schemes and Optimization

Video data analysis through cameras for surveillance has been widely used in areas such as intelligent robots, AR/VR, and smart healthcare. These applications typically use deep neural network technology, which requires significant computational resources. However, terminal devices often have limited computing capabilities, making it difficult to process these tasks. As a result, traditional monitoring video data processing and analysis are mainly uploaded to the cloud for processing. In recent years, with the emergence of edge computing, edge devices have acquired some data processing capabilities, effectively alleviating the long latency caused by limited bandwidth when transferring data to the cloud. At the same time, high precision, low latency response is required for monitoring video analysis to assist various applications. However, due to the limited resources of edge devices and the rapid fluctuations of wireless channels, decisions regarding unloading and resource scheduling become very complex and challenging. This paper mainly studies the computational unloading scheme for real-time video analysis tasks in edge computing systems, focusing on the following:

1. In the case of sufficient edge devices and a small range, such as remote healthcare, the edge-cloud collaborative computing model has become an important way for video surveillance data analysis, which can alleviate problems such as insufficient edge computing power and edge-cloud network communication congestion. When unloading video data, tasks are uploaded to nearby edge devices or the cloud for computation. Under the constraints of long-term average energy consumption and latency, the resolution selection of video, unloading decisions, and edge-cloud communication resource allocation are used to maximize accuracy. By using the Lyapunov algorithm, the stochastic optimization problem is decoupled into independent frame-by-frame optimizations, and a penalty drift function is constructed to minimize the upper bound of the drift plus penalty function to obtain unloading decision solutions.

2. For applications such as security monitoring systems that require large-scale anomaly detection, they need to perform real-time analysis of large amounts of video data obtained by mobile devices. Due to the limited edge-cloud communication resources and long distance from the cloud, the edge-cloud collaborative computing model cannot meet the low latency requirements of massive video data analysis applications. Therefore, this paper unloads the computationally intensive tasks to edge servers near the terminal devices to alleviate the long delay caused by transferring data to the cloud. The dynamic nature of the time-varying channel state and event sequence also significantly affects the decision and analysis task results. When video data is unloaded from multiple terminal devices to multiple edge servers, factors such as video resolution, server selection, bandwidth, and resource allocation affect key indicators such as detection accuracy and real-time task processing success rate. Based on this, this paper considers designing online resolution adaptive and resource allocation schemes to maximize the long-term average utility that balances video analysis accuracy and task processing success rate, and proposes an algorithm based on reinforcement learning to obtain online solutions.