Analysis of Safety-Critical Aspects on Deep Neural Networks and Automotive Dataset for Pedestrian Detection
Deep Learning approaches or Deep Neural Networks (DNN) are being used to address various tasks in image processing such as image classification, localisation, detection, and segmentation. With such signifficant progress in the field of image processing, they become an integral building block when designing pipelines for perception of the environment. Various automotive datasets exist and still many are being developed to enable training of different DNNs, which in turn can be deployed in an Autonomous Vehicle to aid the Advanced Driver Assistance Systems (ADAS) and Highly Automated Driving (HAD) in the vehicles. Various DNNs benchmark themselves against the automotive datasets, to achieve state-of-the-art accuracy. The DNNs detect objects by simply learning to correlate between pixels of the image and representation of an object. Although accuracy serves as a measure of performance, the DNNs are not assessed from a safety critical aspect. For e.g., the dataset might not be an accurate representation of the real world, and hence the DNN might miss the prediction of an object, although test performance was high. Therefore, it is necessary to understand the limitations of the DNN during the perception of environment, especially when dealing with Vulnerable Road Users such as Pedestrians, Cyclists and Motorists. The consequences of failures and insufficiencies in such DNN algorithms are severe and a convincing assurance case that the algorithms meet certain safety requirements is therefore required. As part of this thesis, the connection between safety requirements in automobiles and artificial intelligence is presented. This thesis explores how performance metrics can be used to measure the safety of an object detection algorithm used for pedestrian detection. Additionally proposes new metric to evaluate object detector which can be used to gain insights about the performance and identify challenging scenarios for the object detector.
Heidelberg, Hochschule, Master Thesis, 2021