Supervised learning of risk aware driving rules
Existing work has shown the feasibility of end-2-end deep learning for self-driving vehicles. However, this work lacks the use of metrics with impact on the real world, such as safety. This work is aimed to incorporate safety in the training of an existing Convolutional Neural Network architecture. Safety is incorporated in terms of risk of the current driving scenario. Simple reciprocal of time headway is used as risk metric. This also provides the quantitative measure of risk in training of neural network, making the network risk aware. The first part of the work is the extraction of a dataset from the simulation environment, containing left and right images, scenario information and driving commands. This dataset is then preprocessed to obtain a depth map using the Semi Global Matching algorithm, and risk values using the Risk Metric Calculator. More than 100,000 data points were recorded and filtered based on simple quality metric formulated in this work. The filtered dataset is then used to train a total of six networks, classified into two types, based on their outputs. The networks trained for multiple outputs (steering angle, throttle, risk), are aimed for the supervised learning of risk-aware driving rules. Networks trained for single output (risk), are aimed for dynamic risk assessment using image and depth maps. Such networks can be used as a redundant risk assessment system in vehicles with higher automation levels. The two types of networks are trained and evaluated for three types of inputs, and their results are compared.
Kaiserslautern, TU, Master Thesis, 2018