Training of a perception model on edge of a vehicle

Abstract

An annotation handling system for in an on edge-based manner training a supervised or semi-supervised perception model on edge of a vehicle equipped with an ADS. The annotation handling system stores while the vehicle is being driven, sensor data; selects annotation-eligible data out of the sensor data; generates a learning model candidate by annotating an event in the annotation-eligible data using a perception learning model; generates at least a first corroboration candidate by annotating the event based on perception predictions of the event derived from radar- and/or lidar-based sensor data of the obtained sensor data and/or based on identifying the event in a digital map; determines when one or more of the at least first corroboration candidate match the learning model candidate fulfilling corroboration criteria, an annotation of the event based on the learning model candidate and the first corroboration candidate; and updates the perception model based on the annotation.