A Robust LSTM Model for Video Detection and Classification: An Optimization Procedure

Abstract

The act of human activity recognition plays a vital role in a wide range of fields, such as monitoring and healthcare. Researchers are now using various deep learning and machine learning techniques to analyze and interpret the collected data. With deep learning techniques, such as DL, researchers have been able to improve the performance of various HAR systems by extracting high-level features. In this work, a novel research methodology for human abnormal activities like abuse, assault, arson, arrest, and fighting was introduced. The procedure utilizes the VGG16-based LSTM neural network. The proposed methodology combines the features from the LSTM layers to generate a representation which further supports in enhancing the performance accuracy of the classification task. Due to the complexity of human action recognition and video detection, it can be very expensive to train a model. In this paper, our goal was to minimize the training time and improve the accuracy of our work by implementing a low-cost LSTM structure for video detection. The paper presents a LSTM structure that can perform well after the hyperparameter tuning in validating UCF101's entire dataset. The structure, which is called Context-LSTM, can process the deep temporal features. The use of the proposed LSTM structure can reduce the training time, while maintaining the top-rated accuracy and helps to minimize the memory usage of the GPU. The classification model was able to categorize the "fighting," "arson," "abuse," and "arrest," class labels 95.43%, 95.97%, and 97.37% accuracy, respectively. The proposed model has also tested well and achieved an accuracy rate of 95.81%. with a misclassification error rate of 4.19%.