Self-adaptive SOM-CNN neural system for dynamic object detection in normal and complex scenarios

Abstract

This paper proposes a novel bio-inspired neural system based on Self-organizing Maps (SOMs) and Cellular Neural Networks (CNNs), called SOM-CNN, to detect dynamic objects in normal and complex scenarios. A contribution of our work is a Retinotopic SOM (RESOM) architecture feasible for video and motion analysis. It is inspired by the visual perception mechanism of the human visual cortex, and satisfactorily addresses the disadvantages encountered by other methods in the area. We also propose a new CNN scheme for image thresholding, called Neighbor Threshold CNN (NTCNN), and a self-adapting parameter scheme for the RESOM and the NTCNN models. The proposed system can deal with sudden and gradual illumination changes, dynamic backgrounds, camouflage, camera jitter, and stopped dynamic objects. Experimental results on complex scenarios, using the Precision (Pe), Recall (Rc), F measure, (F1) and Similarity (Si) metrics, yield acceptable average performances with Pe=0.875, Rc=0.8316, F1=0.843 and Si=0.741. Results also show that our proposed system performs better than other methods that have been suggested in the literature. The system can process information at 35fps, rendering it suitable for real-time applications.