Environment perception for an autonomous radio-controlled vehicle with artificial intelligence algorithm

In the last years Autonomous Vehicles have become one of the most important and popular automotive topics. A concept that was considered futuristic a few decades ago is ready to enter our lives and completely change the experience of driving and the whole transportation. Many research studies predict a huge positive impact related to driving aspects such as comfort, low traffic, and safety. In general, an autonomous vehicle’s control consists mainly of three separate modules: environment perception, planning and decision-making, and vehicle control. Environment perception is defined as the process of interpreting vision and sounds. It is a process to interpret, acquire, select, and then organise the sensory information from the physical world to make actions like humans. Therefore, among the technical problems that self-driving vehicles have to address, perception is one of the most challenging. This thesis work was done in collaboration with the student Team Squadra Corse Driverless. According to the race rules, the racetrack boundaries are formed by cones of two colours: blue for the left line and yellow for the right line. For this purpose, the vehicle’s control is not done in a conventional way as lane keeping or similar, but in such a way that vehicle should perceive whether a cone is present and which colour has. In this context, the project work focuses on creating a rapid platform for the PoliTo SC19 electric race-car using a 1/10 radio-controlled vehicle able to test the perception algorithms identified during this project. They are the Single Shot Detector (SSD) and You Only Look Once version 4 (Yolov4) for object detection allowing real-time cone detection. As a first step, a preliminary research on object detection techniques is carried out analysing all the possible Convolutional Neural Networks. Then, the final hardware configuration of the RC car is done wiring all the sensors used for this project that are: stereo camera, Nvidia Jetson Xavier and IMU. In parallel, considerable work is addressed with the software setups using the Robot Operating Systems (ROS) environment. Here, the perception pipeline node performs the object recognition employing either the SSD or the YOLOv4, i.e. it extrapolates the data coming from the stereo camera and gives back the bounding boxes of the cones. Finally, the results and discussion are presented both in a simulation environment and in experimental tests performed in Aeroclub Torino.