Deploying Hybrid Deep Learning Methods In Conjunction With One-Class SVM For Anomaly Detection In Crowded Spaces

The importance of public security and smart city initiatives like the proliferation of monitoring networks and the ability to automatically identify suspicious activity in surveillance footage is growing. Modern methods for recognizing moving objects rely on context modeling. Noise and background movement are introduced because current anomaly detection algorithms focus on geographical information while ignoring temporal information, either entirely or partially. There is a lot of visual and behavioral diversity among the many people in a crowded situation, making it challenging to keep track of individual subjects.To get better and quicker tracking results, a strategy that combines strong segmentation with deep learning is suggested. This method involves taking an extracted video frame and then applying a two-dimensional variance plane and a GLCM to the highlighted pixels inside the video. A hybrid deep learning model is trained using one-class support vector machines using spatial data for robust abnormality detection, and Fuzzy C-Means is used to subsequent frames for segmentation. It assesses recall, precision, and accuracy. When compared to existing approaches, the suggested one achieves better results in experiments when measuring accuracy, precision, and recall. The technique is put into action in MATLAB using the PETS2009 and UCSD datasets.