Adaptive Streaming Continuous Learning System for Video Analytics

Video analytics systems use deep learning models to perform inference on videos and are widely applied in fields such as smart cities and robotics. However, in order to improve inference speed, models with fewer parameters are often chosen to deploy on edge devices, thereby sacrificing the detection accuracy of the task. Continuous learning is an emerging approach to improve the accuracy of lightweight models deployed on the edge for video inference. However, most solutions constantly upload data from edge to the cloud without considering the limited resources and latency of system modules, resulting in the unreliability of the system and loss of model accuracy. This paper investigates the impact of retraining frequency and video encoding configurations on continuous learning. We then design an adaptive streaming continuous learning algorithm (ASCL) with the aim of achieving the desired level of accuracy while minimizing bandwidth resources as much as possible. First, ASCL can adaptively start up according to the needs of users. Second, during the retraining, an adaptive profiling method is designed to select the appropriate encoding configurations to ensure high profiling accuracy. Third, we perform a layer-wise downloading streaming algorithm to ensure secure and smooth transmission. Real-world network traces are driven to the overall evaluation of ASCL. The results of a multitude of videos show the advantages of ASCL over traditional baselines.