Background
Navigating a real autonomous vehicle is a very complex challenge due to the complexity of real world traffic. One of these challenges includes the decision whether to pass or to stay behind leading vehicles. For example, while passing a delivery van that is parked on the side of the road is desirable, passing a car that is waiting in front of a red traffic light should be avoided under all circumstances. The key challenge here is to differentiate between vehicles that are waiting and vehicles that are completely stopped.
Many approaches exist to tackle this static obstacle avoidance problem, however, none of them can cover the full complexity of real world traffic situations . Autoware’s current static obstacle avoidance module1 already addresses parts of this problem: it classifies vehicle-type objects as avoidance targets based on simple heuristics such as lateral offset from the lane center and stopping duration, and excludes objects located near traffic lights, crosswalks, or intersections from avoidance, since these are assumed likely to resume movement. However, this rule-based classification struggles in ambiguous situations and does not generalize well to the wide range of real world traffic scenes. Towards that end, this thesis aims to close this gap by developing a static obstacle detection module in Autoware that will be deployed on EDGAR, our real world autonomous driving platform driving in Munich city center.
Tasks
This thesis includes:
• Conduct a systematic literature review on static obstacle classification and pass/stay-behind decision-making for autonomous vehicles
• Analyze the limitations of Autoware’s rule-based static obstacle avoidance module
• Develop or adapt a method for classifying static obstacles as passable or non-passable
• Integrate the developed module into Autoware’s planning stack
• Evaluate the module in simulation and on EDGAR-collected data sets
• Deploy and test the module on EDGAR in real world traffic in Munich city center
• Document the results in the form of an academic thesis
Contact: marvin.seegert@tum.de