Enhancing Autonomous Vehicle Safety: Using Convolutional Neural Networks For Police Detection

According to the World Health Organization, Ecuador experiences a significant number of road accidents, resulting in numerous deaths, both among vehicle occupants and pedestrians. To address this issue and reduce fatalities, the country is actively working on implementing Autonomous Vehicles (AVs) on its streets. Unfortunately, this technology is far from perfect such is the case of Waymo in which it’s vehicle's failed to respond to a law enforcement officer's command to yield, highlighting the need for improvement in AV technology. To tackle this challenge, we have applied computer vision and artificial intelligence model to accurately identify traffic officers. The methodology consists of using web crawling to collect an image dataset of traffic officers in Cuenca, Ecuador. Then procced to prepare the data set to apply to the models, in this case we used three variants of the YOLO (You Only Look Once) model, specifically YOLOv3s, YOLOv5s, and YOLOv8s and evaluated their behavior. Through experimentation, the YOLOv8s model demonstrated excellent detection capabilities, achieving its an F1 score of 0.78 at a confidence threshold of 0.907. The objective of this model is to enhance AVs' ability to accurately recognize traffic officers, thus improving road safety. As a future enhancement for this project, the researchers plan to create a larger dataset using different images of law enforcement authorities involved in vehicular traffic management in Ecuador. This expansion aims to further improve the model's accuracy and performance.