Video-based human behaviour recognition using hybrid deep learning

Abstract

The increase in the need for effective surveillance in public places has led to the advancement of deep learning techniques for video-based human action recognition. Identifying anomalous human behaviour in video surveillance is essential for public safety, but existing approaches struggle with diverse real-world situations generalization. Most models employ either small datasets or focus on detecting only a few or certain anomalies, which limits their effectiveness. While others overlook the human action recognition essential model performance evaluation metric. This research addresses these gaps by developing a hybrid deep learning algorithm to improve generalization, scalability, and comprehensive human action recognition in surveillance videos. The research proposes an innovative hybrid deep learning model, IncepEffiGuard, integrating InceptionV3 and EfficientNetB7 for feature extraction, and Bidirectional LSTM (BiLSTM) for sequence modelling. Using a random search approach for hyperparameter tuning, the model was trained and evaluated on the UCF Crime dataset, employing Kaggle's computational resources. With an AUC score of 82.12%, the proposed hybrid algorithm outperformed several baseline models. This suggests better generalization and effectiveness in recognizing a range of suspicious human actions across real-world video surveillance situations. The research has achieved its objective of recognizing suspicious human actions using a hybrid deep learning algorithm. The proposed IncepEffiGuard model addresses the limitations and shortcomings of the current approaches, including poor generalisation, dataset constraints, partial anomaly recognition, and limited evaluation metrics. This model provides the basis for real-world applications, improving the safety and security of the public, and ultimately, contributing to SDG 16.