Advanced driver assistance systems (ADAS) are a key technology for improving road safety. Long before fully automated vehicles arrive in significant numbers, ADAS will see high penetration and will substantially reduce accident rates. Supported by Honda, we are developing ways that 5G can "supercharge" ADAS by shedding the limitations of DSRC. The benefits of 5G may then be quantified by demonstrating the benefits of sensor fusion and data sharing in reducing vehicle collision risk.

Sensor Fusion

The Radionavigation lab has built two fully-functioning Sensorium platforms, enabling high-fidelity data collection across multiple sensing modalities. Each Sensorium is equipped with high-definition cameras, radar, and precise GNSS (Global Navigation Satellite System) equipment to collect data about the vehicle's environment. The Sensorium is designed to overcome the challenges associated with arriving at an accurate global positioning solution in challenging environments like Austin's urban core. Of particular interest is the integration of radar-aided localization into our Sensoria. Unlike sensors that rely on visual light, radar is significantly less affected by adverse lighting and weather conditions.

Quantifying Benefits

In order to quantify the benefit of collaborative sensing, sharing sensor data is motivated by a risk assessment problem. In this setting, sensor data from nearby connected vehicles and infrastructure is combined in the cloud into a shared belief state describing static and dynamic obstacles in the vicinity of the ego vehicle. This additional data, collected from different perspectives and modalities, enables the driver of the ego vehicle to effectively see through obstacles. The probabilistic nature of the shared belief state, updated in real time according to Bayesian inference, also identifies risky areas about which little is known.

The utility of the ego vehicle's measurements depends directly on its confidence in its own pose (position and orientation). For this reason, highly accurate data like that provided by the Sensorium platform is critical for a useful risk assessment of the ego vehicle's environment. This framework is also robust to the possibility that inaccurate data is provided by a faulty sensor or malicious attacker. By relying on the consensus of multiple data sources and sensing modalities to collect information about the ego vehicle's "neighborhood", there is ample opportunity to exclude inaccurate data and improbable outliers.

Sensoria Saves Pic