Analysis of environmental conditions on the performances of autonomous driving algorithms

In the last years autonomous driving has been, and continues to be nowadays, one of the most trending topics in the automotive environment. The progress performed in that field is continuously growing, also thanks to a huge amount of investments done by all the main car manufacturers. Vision systems play a key role into this process of driving automation, that’s why during this work two of them will be accurately analyzed through a validation procedure that allows to evaluate their limits and the main critical issues depending on the environmental conditions. The two systems that will be analyzed are: a lane detection system (fundamental to perform autonomous lane keeping) and an object detection system (needed for a lot of functions like the automatic braking or the recognition of road signs). The choice of these two systems is mainly due to the fact that they perform some of the most important tasks needed for autonomous driving. The lane detection system has been implemented in Matlab® using the same working principle of the GOLD system developed by Massimo Bertozzi and Alberto Broggi. The object detection system, instead, is YOLOv3, which is the most powerful algorithm for object detection in real time currently available. It is based on neural networks and has been developed by Joseph Redmond et al. . The validation procedure has been done through a series of tests performed using a specific dataset for each system. Each dataset has been developed taking into account the following parameters: lighting condition, presence of defects and weather conditions. The main aim of this work is to evaluate how and if these parameters influence the level of performance, in order to understand what must be improved and what, instead, has reached a sufficient level of maturity. The results of this analysis shown a certain level of robustness for both the systems, especially for YOLOv3, however also some weaknesses have been discovered. The lane detection system gave bad results in roads with unclear or colored road markings, moreover, bad weather conditions and macro defects leaded to a significant performance degradation. For what concerns YOLOv3, it shown to have a higher level of robustness in every condition, detecting most of the objects even with low resolution input images. The parameter that influenced most significantly the performances are the lighting conditions, which leaded every time to a worsening of the results.