This repository provides the implementation for the paper 'Real-world Anomaly Detection in Surveillance Videos' by Waqas Sultani, Chen Chen, Mubarak Shah.
The project aims to detect anomolous activities in surveillance videos. A pre-trained 3-D convolution network was used to generate input feature vectors and using multiple instance learning an artificial neural network was trained for classification.
UCF-Crime (http://crcv.ucf.edu/cchen/UCF_Crimes.tar.gz) courtesy of Waqas Sultani. It is the original dataset used for the aforementioned paper.
Caffe, Facebook/C3D-1.0 (https://github.com/facebook/C3D), Tensorflow, Python
Resize each video frame to 240*320 pixels and fix frame rate at 30fps.
C3D features for every 16-frame video clip followed by l2 normalization. To obtain features for a video segment, we take the average of all 16-frame clip features within that segment.
We input these features (4096D) to a 3-layer FC neural network. The first FC layer has 512 units followed by 32 units and 1 unit FC layers. Using MIL we try to generate higher anomaly score for anomalous videos than normal videos.
此处可能存在不合适展示的内容,页面不予展示。您可通过相关编辑功能自查并修改。
如您确认内容无涉及 不当用语 / 纯广告导流 / 暴力 / 低俗色情 / 侵权 / 盗版 / 虚假 / 无价值内容或违法国家有关法律法规的内容,可点击提交进行申诉,我们将尽快为您处理。