Direct sunlight causes a minor failure in an optoelectric sensor. Current systems do not respond to this degradation.
Sensing systems have a variety of failure modes, many of which can’t be easily detected. These failures can cause safety-critical events, such as missed obstacles or unnecessary emergency braking. Many systems operate using a “good faith” approach, accepting sensor data at face value. However, to achieve a commercially acceptable level of robustness, a substantially more intelligent approach to data analysis is necessary.
Nominal and degraded operational data is collected for the sensing system to be analyzed. Using a combination of machine learning techniques, this data is analyzed to create an accurate, efficient, description of healthy operation. This description is then used for real-time, on-vehicle analysis to determine the likelihood of the received data being usable. This measure of confidence can be used to intelligently fuse sensors, or provide insight about how a vehicle should respond to incoming data.
Nominal (left) and degraded (center and right) images with confidence values overlaid. Confidence calculations were performed in real time.
Key Elements of the Real-time Sensor Validation
- Provides a measure of sensor health in real time for diagnostics or decision making
- Intelligent configuration selection mechanism ensures optimal performance
- Flexible structure allows for easy integration with multiple sensor types
- Does not require comprehensive failure mode analysis
SwRI has developed a set of tools and procedures that allow for increased awareness of the health of a sensor. The system is capable of detecting and responding to both critical and subtle failures in real time.
- Patent Pending:
Lemmer, S., and Chambers, D., inventors; Southwest Research Institute, assignee. Sensor Data Confidence Estimation Based on Statistical Analysis. United States patent application 14732002, filed June 5, 2015.