随着远程监控和人工智能的融合发展,传统的码率优化算法并不适用于现阶段的移动监控网络场景.在机器视觉应用场景中,相对于传统码率优化算法只关注视频的质量,机器更关注于视频所表达的语义信息.以5G路侧摄像头远程智能检测为应用场景,提出一种基于视频语义的码率优化算法,在有限的码率传输范围内最大化目标检测准确率.具体地,该算法引入视频语义任务模型,将目标检测作为语义任务.分析目标比特与语义之间的特征关系,建立复杂度与运动区域结合的新权重来分配目标比特,使目标检测准确率达到最大化.实验结果表明,相较于HM16.23所使用的帧级树编码单元(Coding Tree Unit,CTU)层码率控制算法,所提算法不仅能够节省码率而且更符合无线远程监控的目标检测需求.在测试环境下平均提升了 1.4％的目标检测准确率,最高能够提升2.5％的目标检测准确率.
Bitrate Control Algorithm Based on Video Semantic
With the integration and development of remote monitoring and artificial intelligence,traditional bitrate optimization algorithms are not suitable for the current stage of mobile monitoring network scenarios.In machine vision applications,machines are more concerned with the semantic information conveyed by videos rather than just the quality of the videos compared with traditional bitrate optimization algorithm.A semantic-based bitrate optimization algorithm is proposed for 5G roadside camera remote intelligent detection as the application scenario,aiming to maximize the object detection accuracy within the limited bitrate transmission range.Specifically,a video semantic task model is introduced to the algorithm and the object detection is treated as a semantic task.By analyzing the relationship between target bits and semantics,a new weight combining complexity and motion regions is established to allocate target bits,thereby maximizing the object detection accuracy.Experimental results show that compared to the frame-level Coding Tree Unit(CTU)-layer bitrate control algorithm used in HM16.23,this algorithm not only saves bitrate but also better meets the object detection requirements of wireless remote monitoring.It achieved an average improvement of 1.4％in object detection accuracy in the test environment,with a maximum improvement of 2.5％in object detection accuracy.
国家科技期刊平台
NSTL
万方数据
