Suyash Dash

Can audio detection improve safety of any vehicle and passengers when object detection fails?

Started Apr. 18, 2024

Abstract or project description

My project is focused on object and audio detection working side by side. With the growing popularity of autonomous vehicle research occurring at the moment, the issue of safety risks in these autonomous vehicles are also prevalent. The issue is that object detection is sometimes not doing its job: during night, dark landscapes, foggy/rainy days, etc. Therefore, I want to incorporate audio detection as a backup to provide extra safety for passengers inside an autonomous vehicle as well as pedestrians on the street. Recently, my dad bought a high end car with only camera detection, assuming its safety features was competent and increased safety for passengers. However, when my dad and I came home through our car, another loud car zooming the road almost crashed with us, which our car could not detect since the other car was coming at our car's blind spot. If there was audio detection that could of detected and localized the other car rushing towards us, our car could of stopped. Luckily, my dad was aware enough to stop when our car's safety function failed, but I was determined to continue my research for safety on the road. Therefore, with object detection working with audio detection, both detection tools can work when the other fails to do so as a backup. If an autonomous car is driving at night and object detection cannot localize nor recognize its surroundings, then audio detection can be programmed to focus with high sensitivity and work for the car until object detection can work. My research can not only help autonomous vehicles, but can work as an added advantage to create an emergency braking system as a way to make any vehicle on the road safer to drive with.

This research explores a novel approach to vehicle safety by integrating audio detection with traditional visual object detection systems. By addressing critical limitations of visual-only systems, especially in low-visibility environments (e.g., fog, rain, and night driving), our hybrid model enhances detection capabilities. Through our approach, we have achieved a 2.25% improvement in detection accuracy compared to purely visual systems. This paper explains the methodology, results, and how distance estimation using audio waveform analysis further enhances safety, comparing this approach with leading technologies like LIDAR and BEVFusion.