SLAM-Based Performance Quantification of Sensing Architectures for Autonomous Vehicles

The list of possible sensors supporting autonomous tasks in a vehicle is considerable. Sensors are one of the main drivers of costs for current prototypes of autonomous vehicles. As production volume increases, companies have to take into account this cost, and consider whether the increase in performance given by the addition of redundancy in sensing is worth the cost. This paper analyzes sensor fusion from a systems perspective, and proposes a method based on factor graphs to quantity the error given by a combination of sensors for the simultaneous localization and mapping task, as a function of the individual sensor capabilities. A total of 81 different combinations are analyzed, with 4 different types of sensors, and varying levels of performance and cost within each type of sensors. This analysis reveals several findings; in some cases, the addition of sensors can decrease the performance of the system by adding noise, and there is a cost over which performance stops increasing. Additionally, we quantity the intuitive idea that systems including sensors working in adverse conditions might still perform well if there are other complementary sensor to provide reliable information to the system.