Improved LinkNet-DenseNet architecture with fine-tuned parameters for abnormal human activity detection for video surveillance

Video-based human abnormal activity detection is to identify and understand abnormal behaviors and actions from the video sequence. For this purpose, an Improved LinkNet-DenseNet (ILN-DN) architecture-based abnormal human activity detection model is introduced in this work, which includes 4 working stages. The novelty of the work lies in the integration of advanced segmentation techniques, enhanced feature extraction methods, and a hybrid classification model for abnormal human activity detection. Additionally, the introduction of a modified optimization algorithm significantly improves convergence speed and detection accuracy. In the pre-processing stage, median filtering is applied to get the pre-processed image (frame), after converting the input video into frames. Afterwards, that pre-processed image gets segmented by the Improved Mask R-CNN model. From the segmented image, features such as Mobile-Based Scale Invariant Feature Transform (MoBSIFT), Improved Shape Local Binary Texture (SLBT), and Hierarchy of Skeleton are extracted. Finally, based on the extracted features, abnormal human activity is detected effectively, with the utilization of a proposed hybrid classification model namely Improved LinkNet-DenseNet (ILN-DN) architecture. The DenseNet architecture is chosen for its efficient feature reuse through dense connections, which enhances the model’s ability to learn complex patterns and improve performance in detecting subtle anomalies. LinkNet, known for its strong segmentation capabilities, is utilized to precisely isolate regions of interest, such as human figures, from the background. Combining both architectures in the ILN-DN model leverages DenseNet’s deep feature extraction and LinkNet’s segmentation precision, making it highly effective for abnormal human activity detection in video surveillance. The parameters of these classifiers are fine-tuned with the utilization of the Improved Red Panda Optimization (IRPO) algorithm, to enhance the detection performance.