SED-NET: Real-Time Suspicious Event Detection via Deep Learning-Based Di-Stream Neural Network

Suspicious event detection (SED) identifies anomalous activities in surveillance data using computer vision and machine learning techniques. However, existing approaches have high false positive rates difficulty distinguishing suspicious from normal behaviors, and limited adaptability to dynamic environments. This research introduces a novel deep learning-based SED-NET model for detecting suspicious events in public places. Initially, input images are collected from two data for detecting Suspicious events. Surveillance camera videos are converted into frames and the suspicious images are pre-processed using a Gaussian adaptive bilateral filter (GABF) to reduce noise while preserving edges. A Sobel edge detector is used to detect the fine edges in the pre-processed frames for enhancing the structural details. Di-Stream Neural Network (DSNN) is introduced with the dual-branch EfficientNet-based feature extractor that retrieves both motion and pose features. The weighted Average Fusion method is used to combine pose and motion features to classify suspicious activities using the Simplified Spiking Neural Network (SSNN). The effectiveness of the proposed SED-NET method was evaluated using specificity, accuracy, sensitivity, and F1 score. The proposed SED-NET model attains an accuracy of 98.97% for UCSD Pedestrian and 98.84% for CUHK Avenue datasets. Moreover, the proposed SED-NET improved the overall accuracy by 7.44%, 4.18%, and 2.73% better than DenseNet121, SegAD, and 3DCNN, respectively.