A Non-Parametric Approach for Modeling Sensor Behavior

Realistic sensor models contribute to the progress of advanced driver assistance systems; off-line development is enabled and rare critical scenarios can be tested. In this paper a non-parametric (i.e., data driven) statistical framework is developed to reproduce sensor behavior. A detailed probability density function is constructed via kernel density estimation by exploiting measurements of an automotive radar system and a high-precision reference system. The approach is capable of inherently modeling sensor range, occlusion, latency, ghost objects, and object loss without explicit programming. Moreover, only few assumptions on the sensor properties are made; therefore, the technique is generic and can be applied to any object-list-generating sensor. The statistically equivalent implementation improvements presented herein render the approach real-time capable. Finally, the method is applied to an automotive radar system using test drives.