Explaining Autonomous Driving by Learning End-to-End Visual Attention

Abstract

Current deep learning based autonomous driving approaches yield impressive results also leading to inproduction deployment in certain controlled scenarios. One of the most popular and fascinating approaches relies on learning vehicle controls directly from data perceived by sensors. This end-to-end learning paradigm can be applied both in classical supervised settings and using reinforcement learning. Nonetheless the main drawback of this approach as also in other learning problems is the lack of explainability. Indeed, a deep network will act as a black-box outputting predictions depending on previously seen driving patterns without giving any feedback on why such decisions were taken.While to obtain optimal performance it is not critical to obtain explainable outputs from a learned agent, especially in such a safety critical field, it is of paramount importance to understand how the network behaves. This is particularly relevant to interpret failures of such systems.In this work we propose to train an imitation learning based agent equipped with an attention model. The attention model allows us to understand what part of the image has been deemed most important. Interestingly, the use of attention also leads to superior performance in a standard benchmark using the CARLA driving simulator.

Vision Transformer for Learning Driving Policies in Complex Multi-Agent Environments

“…Therefore, an attention mechanism can be useful in paying more attention to important vehicles and critical parts of the map in the decision-making problem. Attention mechanism can be used with ConvNets to improve explainability and interpretability of end-to-end deep neural networks [10], [11]. A multi-task attentionaware network with a ConvNet backbone was proposed by Ishihara et al [12] to learn a driving policy via conditional imitation learning.…”

Towards explainable artificial intelligence (XAI) for early anticipation of traffic accidents

“…Kim et al [22] adopted an attention-based method to filter out non-salient image regions to display only regions that causally affect the steering control of a stand-alone vehicle. Similarly, [23] also used an attention model to visualize the perception of deep networks for autonomous driving. Saliency has also been employed to explain AI models for navigation [24], lane change detection [25], and driving behavior reasoning (e.g.…”

“…XAI has been receiving growing attention in autonomous driving. Attempts have been made to explain the functions of various AI models for autonomous driving [21,22,23,24,25,26]. Yet, studies on XAI for AI-powered accident anticipation do not catch the accelerating pace of accident anticipation research.…”

Multi-task Learning with Attention for End-to-end Autonomous Driving

“…In a study [7] conducted by using an open-source driving simulator CARLA [10], it was reported that the driving performance of the imitation learning agent considerably drops under those conditions such as untrained urban area, weather conditions, and traffic congestion. Secondly, it is important to know how well a network perceives visual inputs for such a safety-critical application of autonomous driving, but only a few studies addressed this issue [8,21,24].…”