An Enhanced Real-time Human Detection Keyframe Extraction

Surveillance video assists in criminal investigations by necessitating investigators to dedicate substantial time to scrutinize extensive footage in order to identify suspects. To circumvent this issue, keyframe extraction approach is used to identify the most significant images from these huge stored videos. Currently, there are a number of keyframe extraction methods that aim to identify particular human actions from the frames that have been retrieved. In this paper, the proposed method advances a keyframe extraction approach that identifies keyframes based on human detection to enhance crime scene analysis. The proposed approach consists of two stages. As a preliminary step, the Faster Region Convolutional Neural Network(R-CNN) method is applied to the surveillance video in order to identify any human participants. During the subsequent phase, the video frames that were detected by humans are extracted and refined using the absolute difference method in order to generate a keyframe. Finally, the experiment combining faster R-CNN and keyframe extraction successfully reduced extraction time significantly compared to Histogram of Gradients - Support Vector Machine (HOG-SVM) with background subtraction without sacrificing precision aiding crime investigation.