Robust vehicle classification using sound signals and attention module
Guowei Chi, Zijun Xu, Edmund Sowah, Weiqing Li · International Conference on Smart Transportation and City Engineering 2021 · 2021
Classification of vehicles in urban areas is an important part of transport and traffic management. To improve the performance of car classification, many Intelligent Transport Systems (ITS) have been developed to accomplish this tedious task. Unfortunately, most of these systems are inaccurate in vehicle detection and/or classification. This paper proposes a new car classification algorithm based on vehicle acoustic signals, which adds an attention module to the original BP neural network, highlighting features in the time and frequency domain information that are more capable of distinguishing between vehicle classes, and providing better robustness in real-world environments. To evaluate the performance of the proposed algorithm on the vehicle classification task, we compare the method with the traditional SVM algorithm and the BP neural network without the attention mechanism. Experimental results show that the method using vehicle acoustic features combined with an attention module achieves a classification accuracy of 89.34%, which is a 12.22% improvement over the traditional SVM method, and a 5.23% improvement compared to the BP neural network method without the attention module.