Proposal and Evaluation of a CNN Model Capable of Effectively Handling Long-time Data for Approaching Vehicle Detection Using Sound

Abstract

In Japan, head-on collisions involving automobiles constitute approximately 30% of accidents between vehicles, placing them among the leading causes. Therefore, our research focuses on preventing head-on collisions by studying the detection of approaching vehicles using a Convolutional Neural Network (CNN) based on road environment sounds. To improve the detection accuracy of approaching vehicles using audio data, we believe it is desirable to handle longer-length input data. Therefore, we conducted verification of the impact on detection accuracy by varying the time length of input data to the conventional model. The results indicated that the conventional model may not effectively handle long-length data. Consequently, in this paper, we propose and evaluate a new CNN model that divides the input data at the central time point. As a result, the input data length of 2.49 seconds yielded the highest accuracy, which is 2.49 times longer than the conventional length. Additionally, the detection accuracy of approaching vehicles in the proposed model improved by about 5-10 percentage points compared to the accuracy of the conventional model.