An accuracy–utility-based comparative framework for bandwidth-constrained multi-stream video analytics

Video analytics systems increasingly require adaptive resource allocation to balance analytical accuracy and bandwidth efficiency under constrained network conditions. Utility-based optimization provides a useful framework for representing service satisfaction and guiding allocation decisions across heterogeneous video streams. This paper presents a controlled comparative evaluation of three representative utility-based resource allocation strategies: Alpha fairness-based allocation, Markov chain-based allocation, and Lyapunov drift-plus-penalty (LDPP) optimization in bandwidth-constrained multi-stream video analytics environments. The evaluation is conducted using precomputed video analytics profiling data generated from YOLO-based object detection models under multiple bitrate configurations. Detection accuracy is used as the primary utility input, while bitrate is treated as a resource constraint and reported in bits per second (bps). The study analyzes the relationships among bitrate consumption, detection accuracy, reported utility, and bitrate efficiency. To avoid selection bias, all operating points generated by the three allocators are retained in the full-regime analysis, including extreme Lyapunov operating points. A separate practical-regime analysis is then conducted using a predefined bitrate feasibility threshold to identify deployment-feasible operating points. The results show that Alpha fairness-based allocation provides controlled bitrate usage and predictable allocation behavior under bandwidth constraints. Markov chain-based allocation achieves the highest bitrate efficiency among the evaluated strategies and remains within the practical regime for all evaluated operating points. Lyapunov optimization-based allocation reaches several high-utility operating points, but some of these gains entail excessive bitrate consumption, rendering them impractical under strict bandwidth constraints. These findings indicate that allocator selection should consider not only utility maximization but also bitrate feasibility and bitrate efficiency for practical video analytics deployment.