Abstract
A huge number of cameras records scenes everywhere, generating enormous bulks of videos. Processing these huge masses of videos and detection of abnormal object activities demands adequate resources like time, manpower, and hardware storage, etc. To cope with the aforementioned challenges, our proposed model for an automatic video summarization of abnormal events plays an important role in providing the well-organized storage, quick browsing, and retrieval of the large collection of video data without losing important aspects due to its lightweight. In this research, abnormal object activity detection and summary generation are performed based on two stages i.e. 1) machine learning technique for key event detection, 2) deep learning algorithm to remove extra frames generating summarized video. Firstly, Silhouette images are formed, and two feature descriptors such as Zernike Moments and R-Transform are used to create a combined feature vector. The combined feature vector provides more informative features from images and makes our model lightweight keeping only relevant features. Furthermore, on the combined feature vector, K Nearest Neighbor (KNN) clustering is applied to extract keyframes sequentially. In the end, to improve the performance, Deep Learning Algorithm i.e. ALexNet is trained over preprocessed frames from the dataset. Moreover, the DL classifier aims to eliminate the non-Key Frames and generate surveillance video summaries demonstrating abnormal object activities. The efficiency of the proposed algorithm is analyzed performing an extensive experimentation attaining 99% accuracy approximately.