Application of Genetic Neural Networks to Real-Time Intersection Accident Detection Using Acoustic Signals

Abstract

A genetic neural network (GNN) classification method using acoustic signals is proposed for real-time accident detection at intersections. Backpropagation neural networks (BPNNs) have been widely used in pattern classification. They have fast computation speeds that are desirable for real-time detection systems. However, they tend to converge to local optimums, which consequently affects classification accuracy. The proposed GNN uses a genetic algorithm to improve the global searching ability of BPNNs. GNN performance (i.e., detection rate, false alarm rate, and detection time) is compared with that of a widely used probabilistic neural network (PNN). Test results indicate that the GNN performs comparably to a PNN but with much less computation time. Results of the GNN transferability analysis indicate that the GNN method also is robust and can be successfully applied to intersection accident detection with training data from different sources. Computationally inexpensive and highly accurate, the new GNN is thus suitable for real-time application.