Real-Time Human Detection in World Rally Championship for danger recognition system

This thesis was conducted in collaboration with Marelli S.p.A. Our objective was to develop a network capable of recognizing humans in real-time, intended for use in an alert system in the World Rally Championship (WRC). The WRC is the premier rally competition globally. Despite the high standards of this sport, the safety marshals responsible for overseeing the season's events face challenges in efficiently monitoring all sections of the circuit to ensure spectator safety. In response, Marelli S.p.A. aims to develop an alert system for these marshals, which will help identify individuals in hazardous positions, allowing for timely intervention and enhanced safety across the entire track. A critical preliminary step in developing this system is real-time human detection. Although object detection has been extensively studied, the unique challenges presented by rally racing necessitate additional research specific to this domain. The main challenge was to work with a single camera installed inside the vehicle, which travels at high speeds and needs to recognize people at considerable distances from the vehicle's position. To address this task, we utilized the YOLO (You Only Look Once) network and experimented with various training combinations to optimize its performance for our specific use case and certain challenging situations. We also examined the impact of image quality on the network's performance. After initial evaluation of existing datasets, we decided to create a custom dataset to enhance the network's performance in our domain and ensure robustness against potential scenario changes, despite the limited number of video samples available. The resulting network meets Marelli's standards for accuracy, speed, and performance, having been thoroughly tested on a hardware configuration very similar to the one that will be installed in the car.