Exploiting Redundancy for Reliability Analysis of Sensor Perception in Automated Driving Vehicles

For automated driving, the perception provided by lidar, radar, and camera sensors is safety-critical. Validating sensor perception reliability with standard empirical tests is impractical, owing to the large required test effort and the need for a reference truth to identify sensor errors. To address these challenges, we investigate the possibility of estimating sensor perception reliability without a reference truth. In particular, we propose a framework to learn sensor perception reliability solely by exploiting sensor redundancies. We formulate a likelihood function for redundant binary sensor data without a reference truth and propose a Gaussian copula to model dependent sensor errors. Synthetic numerical experiments show that under an adequate dependence model, correct sensor perception reliabilities can be estimated without a reference truth. Because the selection of an adequate dependence model is challenging without a reference truth, we also investigate how inadequate dependence models influence the estimation. The proposed framework is a step toward the validation of sensor perception reliability because it could enable the learning of reliabilities from a fleet of driver-controlled vehicles equipped with series sensors.