The challenges of processing all this sensing data includes the huge variation in data volume, coupled with the broad spectrum of sampling rates employed. LiDAR can generate more than a million data points per second, while other sensors may be only delivering tens-of-thousands of samples a second. And, while the lower SAE levels can rely upon a handful of sensor inputs, implementation of the higher SAE levels demands significantly more sensor data. Without it, it is not possible to accurately perceive the environment around the entire vehicle.
Fusing of disparate data
In order to achieve an overview of the mass of data being collected it is typically fused together into a comprehensive data stream in a process termed multi-sensor fusion. Through this process incoming sensor data of disparate sampling rate and volume is harmonised to provide a model of the environment that can be used by higher levels of software in the system. This is the approach taken by the SigmaFusion solution from Leti, a technology research institute at the French Alternative Energies and Atomic Energy Commission (CEA). It starts by taking into account that the outputs of range sensors, as used for vehicles, carry some uncertainty. In order to ensure accuracy in the output, and reach safety in the choices made based upon the data, Leti characterises the sensors being used. The source data is fed into the SigmaFusion software which processes it, offering up an occupancy grid to the application layers above. This grid of cells of known dimension assesses the free-spaces and obstacles around the vehicle, with each cell in this model containing the probability of it being occupied or unoccupied by an obstacle.
Within the safety critical system of the vehicle this solution can operate as a stand-alone solution to provide fail-operational perception of the vehicle’s environment for ADAS functions. Alternatively, it is equally appropriate as a safety companion to an automated driving decision system (Figure 1).