Positioning error estimation of honking detection systems based on YOLOX

Traditionally, honking capture system calibration has primarily involved indoor experiments in muffler rooms. Additionally, the manual measurement method for estimating positioning errors in vehicle honking detection systems is not only susceptible to complex backgrounds in acoustic intensity pseudo-color image but also inefficient. Therefore, a value error estimation method based on YOLOX is proposed in this article to estimate the positioning error of the honking detection system. Firstly, the definitions for the horizontal positioning error and the vertical positioning error are provided. Afterwards, CBAM is added in YOLOX network for predicating the honking sound source position and the center of pixel coordinates of pseudo-color images. Based on YOLOX neural network model, a network model for vehicle detection in optical image is trained. Then, the color information from pseudo-color image with vehicles' pixel coordinates is combined to calculate the positioning error. Subsequently, the coordinate system conversion model is established according to the on-site installation information, and the corresponding world coordinates are obtained by combination with the geometric deduction method, to work out positioning error. Ultimately, the outdoor experiment of static and dynamic positioning error of detection systems for vehicle honking is launched. The result indicates that the mean average precision of the improved YOLOX network is up to 98.78%. Besides, the static horizontal positioning error of sound source localization never exceeds 0.15 m, and longitudinal positioning error is 0.31 m. The results show that the positioning error estimation method proposed in this article is helpful for the calibration of the honking detection systems.