An Adaptive Nature-Inspired Paticle Bee Colony Deep Neural Network Technique for Event Detection in Video Sequences

Classifying and detecting difficult video events based on visual modalities remains an uncertain problem. Conventional video presentation approaches ineffectively extract each modality, hindering video event detection (VED) rates. The research methodology involves pre-processing steps: uploading video data samples, extracting video frames, converting to grayscale (rgb2gray), enhancing, and calculating smooth frames from the UCF -101 dataset. This pre-processing phase aims to deliver high-quality frames without data loss. Next, the feature extraction process employs the HoG method to extract feature vectors from refined video frames, facilitating the training process and efficiently reducing dimensions. An adaptive, nature-inspired PBC method is then implemented to select reliable and optimized feature sets from the extracted ones. These selected optimized feature vectors are input into different events of the deep neural network (DNN) classifier. Finally, reliable feature sets are identified through feature matching, and experimental outcomes demonstrate a significant 21.3% enhancement in VED and classification, assessed through accuracy rate, specificity (SP), sensitivity (SN), etc. Compared with traditional approaches using manually designed feature sets, the proposed approach proves more effective. Simulation outcomes on publicly available VED databases consistently outperform state-of-the-art video representation methods such as EFS-linear multi-support vector machine (MSVM), convolutional neural network (CNN), etc.