自动扶梯行为特征监测系统测试数据集

本视频数据采集来自深圳车公庙地铁站。行为算法分析是通过对硬盘录像机实时生成的流媒体文件中每秒截取1-2帧最清晰的图像，首先对截取的图像进行前景检测，识别缺陷区域，在此过程中加入阴影抑制、噪点抑制、画面增强等图像级优化；然后通过图形分割、SVM分类器、神经网络、卷积神经网络等方法对前景进行特征检测，识别出缺陷目标；然后对目标物进行标记，常用的方法有CMT、meanshift、TLD、卡尔曼滤波等，在标记过程中不断更新背景模型，将目标物和背景进行分离，提高检测效果。对于一些检测类的算法如缺陷检测，通过前景检测技术，进行背景分离后即可进入事件检测阶段，自动检测图片中缺陷信息，并完成提醒、统计和标记功能。目前总共采集相关行为算法视频文件50GB。

This video dataset is collected from Chegongmiao Subway Station, Shenzhen. The behavioral algorithm analysis workflow starts by extracting 1 to 2 clearest frames per second from the real-time streaming media files generated by a digital video recorder (DVR). First, foreground detection is applied to the extracted frames to identify defective regions, with image-level optimizations including shadow suppression, noise suppression, and image enhancement integrated into this step. Next, feature detection is carried out on the foreground via techniques such as graph segmentation, support vector machine (SVM) classifiers, neural networks, and convolutional neural networks (CNNs), to recognize defective targets. Subsequently, target tracking and marking are implemented using common approaches like CMT, mean shift, Tracking-Learning-Detection (TLD), and Kalman filtering. During the marking process, the background model is continuously updated to separate targets from the background, thus enhancing detection performance. For detection-oriented algorithms such as defect detection, after background separation via foreground detection technology, the system proceeds to the event detection stage, automatically identifying defect information in images and fulfilling functions including alerting, statistics, and marking. Currently, a total of 50GB of video files related to behavioral algorithm analysis have been collected.

搜集汇总

背景与挑战

背景概述

该数据集是一个用于自动扶梯行为特征监测的测试数据集，包含来自深圳地铁站的视频数据，通过图像处理和机器学习技术识别多种行为特征，如跌倒、大件行李等。数据集总大小为568.06KB，包含2个文件。