Autonomous Vehicle Navigation Using an Integrated Symmetric Multi-Sensor Information Fusion Approach

Autonomous vehicles, which can drive themselves, represent a novel approach to traffic optimization in crowded urban environments. The behaviour of these vehicles is influenced by environmental data, which determine their driving decisions. Consequently, this paper reports an autonomous driving solution, which comprises the specification and implementation of a comprehensive autonomous vehicle driving model. The traffic data are gathered using a multi-sensor information fusion approach, then encrypted and transmitted to the central backend using the public 4G or 5G mobile data network through a symmetric communication channel. The algorithmic model relies on an optimized deep neural network that retains only the necessary features. Validation and training-loss monitoring indicate that the model does not exhibit a diverging train/validation gap. The solution includes advanced 3D object detection primitives. The model was experimentally evaluated using 200 prototype cars supplied by the partner car manufacturer, which supported this research with relevant real-world experimental data. The outcomes of the experimental evaluation process indicate that the system detects relevant road objects in real time, with a detection rate (recall at IoU 0.5) of 98.92%. This figure characterizes the class-correct detection rate only; it does not by itself constitute a safety assessment of the driving system.