Environment perception and road boundary detection from LiDAR data

Self-driving cars are in continuous development and they will give us a safer and most powerful way of moving. The research, in the autonomous vehicles field, is working all the time in order to meet all the requirements in terms of real-time, safety and traffic congestion optimization. To do these, three different development areas are the most crucial: environment perception for mapping and localization, planning and decision-making for the behavioral part and vehicle control in order to give the vehicle the proper input. The environment perception is one of the critical parts because the car has to know what there is in its surroundings: if there something moving like other cars, bikes or people and if there are some fixed obstacles to avoid the collision with them. This is crucial also because the localization of a vehicle cannot be done using only the Global Positioning System or the Inertia Measurements, thus it becomes more and more challenging to have a really good estimate of the mapping environment with high precision. In this thesis work, I proposed a SLAM approach using data coming from the LiDAR sensor. The main strategy is to use a scan-matching algorithm in order to give to the tracker the transformation between consecutive scans, that computes the relative position and then map the entire car surrounding into a single map with the track of the vehicle trajectory. This is done thanks to the Iterative Closest Point algorithm that is a powerful point-set registration process that gives excellent results in terms of accuracy, compared to other methods, for the estimate of the transformation from a moving scan to a reference one. The overall SLAM method has been carried out taking a cue from some other techniques and testing time to time different parameters setups: the car can be larger or smaller, the perceived environment can change based on the road we are on and sometimes we need improved real-time response. The boundary construction problem is strongly related to the environment perception because the car has to know also where it can go without making any damage to things or people, and what the behavioral planner has to command to the controller in order to make a turn or avoid an obstacle. With my work for this part, the vehicle has a free area in which it can move defined, in fact, through the lines of the boundaries obtained thanks to an interpolation method. The idea is to discretize the processed scan into shapes and then obtain the front free space of the vehicle. This boundary construction process is entirely made through point cloud processing and shapes analyzing. The overall coding and algorithms have been developed using MATLAB and related toolbox and they have been tested with some experimental data took with a Velodyne VLP-16 LiDAR.