Condensation of Video for Activity Detection in CCTV Footage using Deep Learning

Video Surveillance is a very critical aspect from a security perspective for any organization or a residential compound. Video surveillance involves capturing video feed from multiple closed circuit television cameras. It is also preferable to have the video data in high resolution to better understand any incident of interest that is captured in the video footage. One of the major challenges with storing high resolution video footage is the volume of data available and the data storage space required. Another challenge posed to the users is the time spent in reviewing the video footage to identify the video segment of interest in order to identify an incident. Deep learning algorithms like Convolution Neural Network is an effective solution to recognize image data classification. Convolutional neural networks are utilized in this research to classify individual frames of video data to recognize if there is any activity in the frame or not. CCTV footage from a classroom was used to train and validate the proposed model. A validation accuracy of 99% was observed. The sequence of frames with activity are recognized and only those frames are combined together into a video file. The resulting video file is a condensed video which only shows the video segments with the activity that the model is trained to recognize.