Deep Learning for Autonomous Vehicle Real-time Hazard Detection and Avoidance

Abstract

[1] [2]Using real-time sensors and human-machine interfaces, autonomous vehicles have the capability to process various sets of input data to learn and respond to their environment. Artificial neural networks (ANN), depending on its architecture, perceive, learn and make decisions, given the learning dataset and deployment scenarios. Despite their significant capabilities, ANN models do no guarantee flawless predictions or decision making and their mechanisms remain limited and are ambiguous, some argue that. Artificial intelligence (AI) models may lack transparency and understanding of their decision making, resulting in accidents or legal consequences in unsupervised deployment scenarios. Deep learning (DL) models are especially criticized for their black-box nature and potential vulnerabilities. Development of AI architectures with higher transparency and better learning and decision-making capabilities in autonomous vehicles have become a necessity for minimizing potential false predictions and learning optimizations. As a novelty detection tool, saliency maps in neural networks, have been widely used in deep learning for detection and to highlight input structures contributing most to outputs.[3]In the context of deep learning, the system decides on the strategies to pass and reflect different road representations (like pedestrians, traffic objects, road signs) to the driver via a camera system. The backend of the DL system has to perform 2 classification jobs (a) Path Planning, (b) Hazard Avoidance. Online performance issues of the system include the heavy computational power requirement, and ambiguity of the decisions. Thus, the FPD-based system converges much faster than the MLP and achieves the best precision amongst the DL based strategies. The DL based poll is initially the most ambiguous strategy but becomes the most confident one with increasing sample size. In the back-end, the differences in classification performance from different DL-based architectures are negligible, which can be interpreted as independence from changes in classification architectures. The impact of the path planning policies in the hybrid system is much more drastic compared to the DRFranja classification strategies. In the simulated training environments, the different vehicle types pose more ambiguity and DL-based frontends are faster and more accurate in learning on their hazard perception than ML-Perceptron strategies.