Occlusion handling in behavior planning using imitation learning for autonomous driving

Abstract

Commissioning a self driving vehicle to run on road, requires the facilitation of complete vehicle system to work at all conditions. Behavior planning is a crucial part of the autonomous driving system and it is important to ensure safe and comfortable navigation of the ego vehicle. More advancements are required to enhance the data-driven approaches for the planning systems. The urban driving scenarios always possess a variety of disturbances and inefficiencies. In which, the roundabout is a challenging driving task where uncertainties are caused due to static priority rules and occlusions that limits the field of view for the ego vehicle. Thus behavior planning must make sure to consider the uncertainty of limited visibility of the environment explicitly. Although machine learning-based approaches show promising results for behavior planning. A single planner cannot handle all other urban driving scenarios. Hence, an imitation learning-based technique can help the behavior planner to mimic the human expert behavior. In this context, an end-to-end planning system based on imitation learning proposed by Waymo is used. The behavior planning framework makes use of mid-level input and output representations making it viable to be interfaced with existing vehicle system. The planner outputs a set of waypoints to drive the vehicle controller. However, the existing imitation learning-based planning framework with the Intelligent Driver Model (IDM) as an expert and policy model made of a multi-task network did not address this use case of occluded roundabouts. As the default IDM generates training data with a visibility of the environment, there arises a need for a strategic approach to handle the occluded environments. This thesis work aims at leveraging the existing planning system to handle the situations in a roundabout with limited visibility. Ultimately, the goal is to train the policy model with more realistic data and enable it to make safe and comfortable driving decisions. For this purpose, an occlusion algorithm is implemented to induce limited visibility of the roundabout environment in simulation. And the expert model is enhanced to handle the limited field of view much similar to how a human driver behaves. Consequently, the training dataset generated from the expert is upgraded with an additional input feature. This add-on feature in the input data provides enough knowledge for the policy to perform well in the occluded environment. A study on modern architecture search is performed and a suitable convolutional network is adopted as the backbone for this multi-task model. The enhanced behavior of the proposed approach is demonstrated via detailed quantitative analysis. For this purpose, a new comfort metric is defined and used as Key performance Indicator (KPI) to evaluate the models. An ablation study is conducted with the expert and confirmed that the new extended IDM behaves more carefully in an occlusion environment. In the end, the influence of the training data is inferred by a detailed comparison of the policy model in default and occlusion environments with different dataset configurations. The importance of more realistic data is realized and also shows that the policy model can imitate the expert behavior well enough. It is exhibited that the proposed methodology can handle the occlusions in the complex roundabout situations in simulation.