Improved convolutional neural network and spectrogram image feature for traffic sound event classification
In recent years, Sound Event Classification (SEC) is a remarkable field of computer audition research. The abilities to automatically capture and classify sound events are meaningful for expanding the auditory ability of autonomous vehicle in a complex driving environment. The Autonomous vehicles will encounter some special road conditions during driving, such as large truck occlusion, tunnels, and buildings that hinder radar detection. Therefore, the perception capabilities of on-board lidar, vision, and other sensors are impacted, resulting in a great threat to the safe driving of the car. At this time, auditory environment perception can provide auditory dimension information for the vehicle. However, with the increase of noise level, the SEC task of autonomous vehicle becomes seriously difficult. Traditional machine learning methods show weak robustness in the highly noisy environment. In this work, an autonomous vehicle Sound Event Classification (AVSEC) framework based on Spectrogram Image Features (SIFs) and Convolutional Neural Networks (CNN) algorithm is proposed to solve the SEC question. Further on, the ambulance, bus, civil defense siren, fire truck, screaming, etc. are taken as the research traffic sounds. Traffic Sound Datasets (TSD) are collected, augmented and customized. Then, directly connected channel (DCC) between convolutional blocks and Convolutional Block Attention Module (CBAM) are designed to optimize the neural network. Testing results show that the AVSEC system achieves an accuracy of 97.18%. Furthermore, the results of AVSEC and other machine learning methods are compared, which shows that the combination of SIFs and AVSEC-net provides better noise robustness and classification accuracy in AVSEC task. The sound classification algorithm designed in this paper, as an advanced auxiliary driving system for sound detection, can identify the sound types in the current traffic environment of the Autonomous vehicles, that is, the traffic scene category. The vehicle computer can remind the driver to pay attention to driving safety through the above sound types, so as to improve the safety of the driver and the vehicle.
- Record URL:
-
Availability:
- Find a library where document is available. Order URL: http://worldcat.org/issn/09544070
-
Supplemental Notes:
- © IMechE 2023.
-
Authors:
- Xu, Ke
-
0000-0003-2200-905X
- Yao, Jingyi
- Yao, Lingyun
- Publication Date: 2024-11
Language
- English
Media Info
- Media Type: Web
- Features: References;
- Pagination: pp 4230-4244
-
Serial:
- Proceedings of the Institution of Mechanical Engineers, Part D: Journal of Automobile Engineering
- Volume: 238
- Issue Number: 13
- Publisher: Sage Publications Limited
- ISSN: 0954-4070
- EISSN: 2041-2991
- Serial URL: http://pid.sagepub.com/content/current
Subject/Index Terms
- TRT Terms: Autonomous vehicles; Neural networks; Sound; Spectrographic analysis; Traffic noise
- Subject Areas: Environment; Highways; Vehicles and Equipment;
Filing Info
- Accession Number: 01937396
- Record Type: Publication
- Files: TRIS
- Created Date: Nov 19 2024 2:38PM