Audio-Based Vehicle Detection System: Enhancing Safety for Bicycle Riders

Abstract

To enhance the safety of bicycle riders an android mobile application was developed that utilizes real-time audio analysis to detect approaching vehicles and alert the riders when there is a safety issue. The application employs a CNN lite model to detect the presence of the vehicle. The CNN model was trained on a dataset of 0.5-second audio clips of vehicle and non-vehicle sounds. The root mean square value of the audio clip is calculated to figure out the approaching vehicles. Based on the model prediction and the root mean square value of amplitude, the application issues alerts to the user. Three alert levels are defined, ranging from level one for low amplitude to level three for high amplitude. The alert types are audio, visual and vibration. Users can customize and adjust the alert types and threshold values within the predetermined range according to their preferences. The evaluation of the model revealed favourable results, as indicated by low loss and high recall and precision metrics, affirming the efficacy of the model for accurately detecting vehicles. The ability of the model to proactively detect and alert bicycle riders of incoming vehicles, enable the users to take timely actions to prevent potential hazards and positions the application as a crucial and effective lifesaving tool.