A Method for Learning Automated Driving Systems and Data Consistency Analysis

Abstract

In this thesis a method for developing a deep learning network to predict signals in automated driving systems is presented. The goal is to incorporate the network in a tool aimed to find discrepancies in vehicle testing data. To this end, a neural network was constructed and trained on simulations of vehicle dynamics with the purpose of predicting the dynamics of a real test vehicle. A parallel goal was to compare the properties of the network to a software for simulating vehicle dynamics. The thesis was made in collaboration with Zenseact, an AI and software company developing the complete software stack for automated driving and advanced driver-assistance systems (ADS and ADAS). The network did predictions replicating a simulation software’s Controllers, Vehicle model and the two combined. To evaluate which type of training data gave the best result, three datasets with different driving schemes were generated by the simulation software. On a test set constructed from 21 log files from real test drives, the network got the best results predicting the Controllers and Vehicle model combined. In this case, for only predicting a vehicle’s lateral motion, a training set consisting of constant longitudinal velocity generated the best results. If both lateral and longitudinal quantities are desired, a dataset with sinusoidally changing acceleration performed best overall. The other purpose was to compare the time aspects of the network to the simulation software. This was done by evaluating the time for data generation, training and prediction on the ”Constant” dataset used in this thesis. This dataset is comprised by 84,528 time steps of ADS signals. Data generation took 8.5 to 9 minutes over all tests. The training time was consistently 61 to 62 seconds in total and the prediction time on the 84,495 batches was between 6 to 7 seconds, or an average of 75.9 μs per batch. It was concluded that, apart for the intended goal of being a core in a data consistency tool, the main use case for the network could be in applications where fast data generation times are crucial, for example in near real time predictions during test drives.