Abstract: Video data provides ample information for intelligent networking of vehicles.To improve the quality of video content extraction and keep more valid information in offloaded videos, this paper presents a content-driven method for instructing computation offloading, and a decision-making algorithm based on improved Monte Carlo Tree Search(MCTS) under delay constraints.The video content is pre-processed through key frame extraction at the vehicle end, which helps to analyze the importance of content analysis tasks, making important tasks obtain more computation resources.Then a heuristic search algorithm based on reinforcement learning is adopted for offloading decision making, and a DNN is adopted to pre-train the priori transition probability.So the convergence speed of the algorithm can be optimized and the computational complexity is reduced.Experimental results show that the proposed algorithm can effectively reduce energy consumption and improve the accuracy of video content understanding under the delay constraints.Compared with the algorithms based on Q-learning and simulated annealing, the proposed algorithm displays a higher convergence speed and lower computational complexity.Its overall system utility reaches 37% when constrained by a delay of 700 ms.

Key words: Monte Carlo Tree Search(MCTS), video content understanding, computing offloading decision, edge computing, Internet of vehicles

摘要： 视频数据能够为车辆的智能网联化提供丰富的信息，为了更好地提取视频内容并使卸载后的视频中包含更多的有效信息，在时延约束条件下，设计一种内容驱动的计算卸载指导方式并提出基于改进蒙特卡洛树搜索的计算卸载决策算法。在车辆端通过关键帧提取来对视频内容进行预处理，以有效分析视频内容理解任务的重要性，使得更重要的任务能够获得更多的计算资源。采用基于强化学习的启发式搜索算法完成计算卸载决策，并引入深度神经网络预训练先验转移概率，从而优化算法的收敛速度并降低计算复杂度。实验结果表明，该算法能够在时延约束下有效降低能耗并提升视频内容理解精度，相比基于Q-learning、基于模拟退火的算法，其收敛速度更快，计算复杂度更低，在700 ms时延约束下系统总效用达到37%。

关键词: 蒙特卡洛树搜索, 视频内容理解, 计算卸载决策, 边缘计算, 车联网

CLC Number: 

• TN992.52