| Abstrakt: |
There is an increasing demand for automated violence detection with a wide range of threats in society and less manpower to monitor them. Especially, detecting violence in crowded scenes is challenging because of the rapid movement, overlapping features due to occlusion, and cluttered backgrounds. This paper plans to implement the enhanced model for video violence detection with the aid of intelligent approaches. The proposed model covers different phases like (a) pre-processing, (b) feature extraction, (c) optimal feature selection, and (d) classification. Initially, the video frames are split, and the pre-processing of the frames is carried out by the Gaussian filter. Next, the feature extraction procedure is undergone, in which the Motion Boundary Scale Invariant Feature Transform (MoBSIFT), Histogram of oriented Gradients (HoG), and Motion Weber Local Descriptor (MoWLD) are used. Further, the optimal feature selection is adopted. The hybridization of two well-performing algorithms like Spider Monkey Optimization (SMO), and Grasshopper Optimisation Algorithm (GOA), namely Spider Monkey-Grasshopper Optimization algorithm (SM-GOA) is used for optimal feature selection with the intention of solving a multi-objective function. Then, the classification of violence and non-violence video frames is done by the Deep Neural Network (DNN), in which the training algorithm is enhanced by the same SM-GOA. The proposed SM-GOA-based DNN had achieved less False Positive Rate (FPR), False Negative Rate (FNR), and False Detection Rate (FDR) values compared to the existing methods proving less variance to the negative performance and thus has shown good overall performance on the violence flow dataset. Experimental results on diverse benchmark datasets have demonstrated the superior performance of the proposed approach over the state-of-the-art. [ABSTRACT FROM AUTHOR] |
Full text from SpringerLink