Communication University of China
North China Institute of Science and Technology

随着人工智能的发展,视频识别技术取得了长足进步。然而,在许多应用场景中,获取的视频存在清晰度低、物体遮挡、场景复杂以及烟雾遮蔽等问题,使得智能模型的识别精度与速度下降。本文针对人体动作或行为识别、场景识别以及情感识别三个经典视频识别任务,分别就引入多源信息与基于视频单源信息两方面总结面向复杂环境视频的识别研究方法。多源信息主要介绍了引入与低质量视频同一环境下同一时间获取的识别目标的其他模态信息进行辅助识别的方法,这类方法通常将多元特征在隐式空间中对齐,以期获得特定于任务的联合表征,充分发挥多源信息的互补特性。单源信息主要介绍仅依靠视频自身时空信息或语义综合的方法,这类方法通常深度挖掘视频内容的空间语义特性以及时间编码特性,使得目标信息被凸显。

With the development of artificial intelligence, video recognition technology has made great progress.But, in many application scenarios, the acquired video has problems such as low clarity, object occlusion, complex scene and smoke occlusion, which makes the recognition accuracy and speed of the intelligentmodel decrease.Aiming at the three classic video recognition tasks of human action or behavior recognition,scene recognition and emotion recognition, this paper summarizes the research methods for video recognition incomplex environments from two aspects of introducing multi-source information and based on video single-source information.Multi-source information mainly introduces the method of auxiliary recognition by introducing other modal information obtained at the same time in the same environment as low quality video.Thesetype of methods usually align multiple features in implicit space to obtain task specific joint representations,57fully leveraging the complementary characteristic of multi-source information.Single-source information mainlyintroduces the methods that only rely on the spatio-temporal information of the video itself or semantic synthesis.These methods typically deeply mine the spatial semantic and temporal encoding characteristics of videocontent, so that the target information is highlighted.

张竣淞,汪洋.面向复杂环境的视频识别方法研究[J].华北科技学院学报,2023,20(4):75-81

ZHANG Junsong,WANG Yang. Research on video recognition method for complex environment[J]. Journal ofNorth China Institute of Science and Technology,2023,20(4):75-81