Application of Neural Networks for Validation of Data Integrity in Large Automotive Radar Datasets

The automotive industry, particularly Active Safety and Autonomous Driving, requires a vast amount of data collected during test drives for machine learning and system validation. The data’s integrity is necessary to ensure the correct operation of the algorithms. In a situation where the amount of collected material exceeds thousands of hours, it becomes necessary to use automatic verification methods. Data integrity can be ensured by comparing signals with estimations based on measurements independent of the verified value. It often turns out, however, that designing such an algorithm and its accurate calibration requires considerable effort. This work proposes using a neural network as an alternative to an algorithmic approach to reduce development time drastically. The applied solution allows the estimation of the vehicle’s speed and yaw rate based on the detections from a single radar scan with a neural network consisting of less than 50,000 parameters. The accuracy of the estimation is sufficient to detect anomalies in the data by comparing the measured and estimated values. Further, a neural network that allows the direct detection of anomalies in the input data is presented.