Title: Anomaly object detection from video with hybrid activation function-enabled optimised deep convolutional neural network

Authors: Sangita Mahendra Rajput; M.D. Nikose

Addresses: Department of Electrical and Electronics Engineering, Sandip University, Nashik, Maharashtra, 422213, India ' Department of Electrical and Electronics Engineering, Sandip University, Nashik, Maharashtra, 422213, India

Abstract: The research proposed a novel DL model named Interactive Routing Algorithm Optimised Hybrid activation function enabled channel-wise deep convolutional neural network (IRA-C2NN) for accurate video anomaly detection. The model enhances the detection accuracy concerning the key parameters of the video input in hybridisation with the activation function. The selected frame from the video is subjected to moving object detection and tracking, the identified anomaly object is tracked in every frame using the YOLOv5. The combination of the pigeon and the particle swarm creates the IRA optimisation, which is hybridised with the classifier to enhance the detection rate of the anomalies, which minimises the error. The achievement of the research is shown as 96.03% with accuracy, 97% with sensitivity, 94.32% with specificity, and 96.33% with the F1 score.

Keywords: anomaly detection; video surveillance; key frame selection; deep learning; convolutional neural network.

DOI: 10.1504/IJSISE.2025.151429

International Journal of Signal and Imaging Systems Engineering, 2025 Vol.14 No.2, pp.124 - 146

Received: 31 Jan 2025
Accepted: 26 Sep 2025
Published online: 29 Jan 2026 *

Full-text access for editors Full-text access for subscribers Purchase this article Comment on this article