Real-Time Violent Action Recognition Using Key Frames Extraction and Deep Learning

Violence recognition is crucial because of its applications in activities related to security and law enforcement. Existing semi-automated systems have issues such as tedious manual surveillances, which causes human errors makes these less effective. Several approaches been proposed using trajectory-based, non-object-centric, deep-learning-based methods. Previous studies shown that deep learning techniques attain higher accuracy lower error rates than those other However, the their performance must be improved. This study explores state-of-the-art architecture convolutional neural networks (CNNs) inception V4 detect recognize violence video data. In framework, keyframe extraction technique eliminates duplicate consecutive frames. keyframing phase reduces training data size hence decreases computational cost by avoiding For feature selection classification tasks, applied sequential CNN uses one kernel size, whereas v4 multiple kernels for different layers architecture. empirical analysis, four widely used standard datasets are with diverse activities. The results confirm approach attains 98% accuracy, cost, outperforms existing detection recognition.