【作者】 左萌；

【导师】 王敬宇；

【作者基本信息】 北京邮电大学 ， 计算机技术（专业学位）， 2020， 硕士

【摘要】 随着计算机视觉技术的不断发展,视频类产品也逐渐走向智能化。人工智能算法通常需要大量样本数据进行训练,由此大量开源数据集应运而生。然而,在将人工智能算法应用于实际工程时,由于待处理问题具有不同的特征,仅使用公共数据集是不够的,还需要私有数据集更有针对性地训练。单纯使用人工标注的方式构建私有数据集,需要消耗大量的人力和时间成本。而现有的一些利用众包或抽样进行标注的系统又存在标注过程复杂、准确率低等缺点。为解决上述问题,本文设计并实现了一种基于目标检测与跟踪的辅助标注系统,帮助人工进行视频的半自动标注,降低了人工标注的工作量和难度,大幅提升标注效率。首先,本文对目标检测和目标跟踪算法进行研究。应用FasterR-CNN、YOLOv3和SSD三种目标检测算法分别与卡尔曼滤波跟踪算法进行组合,在每一帧视频中生成目标检测框和目标跟踪框,并使用匈牙利算法进行目标指派,生成该帧目标对应的标注信息。用户可根据视频选择合适的算法进行初始标注,以提高标注的精准度。其次,对关键帧进行提取,进行人工校验。本文采用镜头边界检测算法和自定义易错帧选择算法在标注算法计算的过程中提取出关键帧,并定义人工校验规则,结合提取的关键帧进行修正后,再重新初始化辅助标注算法,从而实现对整个视频的修正。最后,构建辅助标注系统,该系统实现对用户上传的视频进行辅助标注的功能。系统通过实现辅助标注算法,对视频进行初始标注,然后通过人工校验的方式修改关键帧中目标框的位置、大小等信息。标注完成后,用户可下载所需格式的标注文件。经测试表明,本系统能够对视频标注过程起到良好的辅助作用,大幅缩短了标注时间,具有很高的实用性。更多还原

【Abstract】 With the continuous development of computer vision technology,video products have gradually become intelligent.Artificial intelligence algorithms usually require a large number of sample data for training,so a large number of open source data sets appear.However,when applying artificial intelligence algorithms to actual projects,because the problems to be processed have different characteristics,it is not enough to use public data sets,and private data sets also need to be more targeted training.Constructing private data sets using only manual annotation requires a lot of manpower and time costs.However,some existing systems that use crowdsourcing or sampling for labeling have the disadvantages of complex labeling process and low accuracy.In order to solve the above problems,this thesis designs and implements an auxiliary tagging system based on target detection and tracking,which helps manual semi-automatic tagging of videos,reduces the workload and difficulty of manual tagging,and greatly improves the efficiency of tagging.First,this thesis studies the target detection and target tracking algorithms.The three target detection algorithms of Faster R-CNN,YOLOv3 and SSD are combined with the Kalman filter tracking algorithm respectively to generate the target detection frame and target tracking frame in each frame of video,and use the Hungarian algorithm for target assignment to generate the frame The label information corresponding to the target.The user can select a suitable algorithm for initial labeling according to the video to improve the accuracy of labeling.Secondly,extract key frames and perform manual verification.In this thesis,the lens boundary detection algorithm and the custom error-prone frame selection algorithm are used to extract key frames during the calculation of the labeling algorithm,and define manual verification rules.Implement corrections to the entire video.Finally,build an auxiliary annotation system,which implements the auxiliary annotation function for the videos uploaded by users.The system realizes the initial annotation of the video by implementing the auxiliary annotation algorithm,and then modifies the information such as the position and size of the target frame in the key frame through manual verification.After the annotation is completed,the user can download the annotation file in the required format.Tests show that the system can play a good auxiliary role in the video labeling process,greatly shorten the labeling time,and has high practicality.更多还原