安防系统人形检测数据
收藏资源简介：
用于检测图片内出现的人形目标，提供人形目标的位置信息（坐标框），可以直接应用于各类场景，如：家用安防相机的动检事件过滤，对动检事件上报的图片进行人形检测，过滤不存在人形的图片以及在存在人形图片上标记出存在的人形目标供用户查看从云端存储服务器接受智能摄像头动检时间视频片段，基于FFMPEG音视频框架对视频片段进行视频流解码，生成输入图片信息；将生产的输入图片信息按序进行预处理，包括像素归一化与图像缩放，将预处理后的图像通过PPLCNet分类网络进行人形识别判断，输入过滤阈值：当图像中目标占比超过20%时输出为有人，否则为无人；将有人图片帧进行标记形成图片集合切片【i,j】送入人形检测缩略图生成阶段，输入过滤阈值并通过yolov6-s模型进行人形检测网络生成人形检测结果；根据网络的人形检测结果在图片上进行绘制，将绘制的图片集合合并生产webp缩略图并输出位置信息。
This dataset is developed for detecting human targets in images and providing their bounding box position information, which can be directly deployed in various scenarios. For instance, it can be used for filtering motion detection events of home security cameras: conducting human detection on images reported by motion detection events, filtering out images without human presence, and marking detected human targets on images with humans for user review. The processing pipeline first retrieves video clips of motion detection events from smart cameras via cloud storage servers. Subsequently, it decodes the video streams using the FFMPEG audio-video framework to generate input image data. The generated input image data undergoes sequential preprocessing, including pixel normalization and image resizing. The preprocessed images are then fed into the PPLCNet classification network for human presence recognition and classification, with a predefined filtering threshold: when the proportion of human targets in an image exceeds 20%, the model outputs "human present"; otherwise, it outputs "no human present". The image frames identified as containing humans are marked to form image set slices [i,j], which are input into the human detection thumbnail generation stage. The same filtering threshold is applied, and the YOLOv6-s model is utilized as the human detection network to generate human detection results. Finally, bounding boxes are drawn on the images based on the detection results, the drawn image sets are combined to generate WebP thumbnails, and the position information of the detected human targets is output.
