Enhancing public safety: a hybrid Conv_Trans-OptBiSVM approach for real-time abnormal behavior detection in crowded environments

Enhancing public safety: a hybrid Conv_Trans-OptBiSVM approach for real-time abnormal behavior detection in crowded environments

연구 분야: Verification

학회: Signal, Image and Video Processing

Abnormal behavior identification becomes significant in real-time smart environments as the act of threats is increasing globally, nowadays. Accurate recognition of abnormal behaviors well-ensures public safety and security, especially in crowded scenes, but is more complicated to estimate. The closed-circuit televisions (CCTVs) installed in public places to prevent crimes demand automated behavior modeling mechanisms to detect abnormal activities. The deep learning (DL) based computer vision algorithms although performing very well, are not capable of detecting abnormal behaviors in CCTV images in real-time due to the high computational complexity and ineffective learning behavior. To overcome this limitation, in our research, an intelligent ‘Hybrid Conv_Trans-OptBiSVM’ based abnormal behavior detection model is proposed. Diverse model components such as convolution backbone layer, spatial–temporal encoder and Attention in attention mechanism (A2M) are integrated for extracting complicated data patterns to identify the abnormal events in the image frames. The 2D-CNN layer extracts local and high-level features from the images. The encoder layer aims to identify global space and long-range temporal dependencies among adjacent pixels using self and cross-attention with temporal association. In addition, an A2M method assists in enhancing the quality of correlation map. It searches for correlation uniformity surrounding every key to improve the relevant correlations of corresponding key query pairs. Finally, classification is done by the designed optimized binary support vector machine (OptBiSVM). It uses particle swarm optimization (PSO) algorithm for tuning hyperparameters such as kernel parameter and cost parameter. We compare our model’s performance with other algorithms to evaluate and validate its effectiveness using multiple benchmark datasets- UNM, UCSD (PED 1, PED2, and PETS 2009. The notable outcomes generated by the Hybrid Conv_Trans-OptBiSVM algorithm emphasize supreme anomaly detection performance. Thus, these findings demonstrate our research’s robustness for recognizing and monitoring abnormal events in crowded scenes.