Highly automated driving offers the potential to make traffic significantly safer for all road users. In this manner, it will have a sustainable influence on the lives of people. To turn this
vision into reality, big challenges need to be solved on the technical level. To enable a car to find and plan its own way through the traffic in a safe and autonomous way requires first of all a robust and accurate sensing of the environment. This challenge can only be solved by using different and redundant sensor systems that are intelligently processed and fused. Both research and industry agree that Radar, Lidar, and cameras are the required sensor technologies. However, the questions of the optimal processing of the data and of the architecture of the processing units are unsolved.
Sensors within an autonomous car have to take over the responsibilities of the human perception: Observation and evaluation of the current traffic situation as well as detection of
road users and objects that potentially interact with the trajectory of the car itself. The sensors in the car however sense the environment continuously and in all viewing directions.
This causes huge amounts of raw data which require an intelligent classification of the relevant data, comprising the classification of the objects as well as the detection and
mitigation of interference signals.
To answer these questions, a hardware accelerator for binary Bayesian neural networks (BBNNs) the project SAHaRA will be developed. Deep neural networks (NNs) are the most
successful machine learning algorithms as of today and over the last years deliver the best classification results for different problems in the literature.
Within SAHaRA, the advantages of BBNNs with respect to conventional deep NNs shall be investigated and shown with the application of autonomous driving. To this end, the
developed accelerator will be integrated in a smart radar sensor, which is a sensor with integrated intelligent signal processing. The following shall be shown within SAHaRA:
- BBNNs can be implemented in hardware highly efficiently with respect to computational and energy consumption requirements.
- By employing Bayesian learning algorithms, BBNNs are able to deliver statistical uncertainties about the classification results. This can potentially improve a robust data fusion in the car significantly.
- BBNNs within a radar sensor allow for a detection of relevant data directly in the sensor. This relaxes requirements on the central processing unit in the car and optimizes the partitioning of the computations as well as the energy efficiency.
The theoretical questions about the required architecture and the learning algorithms of the BBNNs will be enabled by data acquired within the ALP.Lab testing region.