5 October 2021 in IEEE Access
Autonomous vehicle (AV) industry has evolved rapidly during the past decade. Research and development in each sub-module (perception, state estimation, motion planning etc.) of AVs has seen a boost, both on the hardware (variety of new sensors) and the software sides (state-of-the-art algorithms). With recent advancements in achieving real-time performance using onboard computational hardware on an ego vehicle, one of the major challenges that AV industry faces today is modelling behaviour and predicting future intentions of road users. To make a self-driving car reason and execute the safest motion plan, it should be able to understand its interactions with other road users. Modelling such behaviour is not trivial and involves various factors e.g. demographics, number of traffic participants, environmental conditions, traffic rules, contextual cues etc. This comprehensive review summarizes the related literature. Specifically, we identify and classify motion prediction literature for two road user classes i.e. pedestrians and vehicles. The taxonomy proposed in this review gives a unified generic overview of the pedestrian and vehicle motion prediction literature and is built on three dimensions i.e. motion modelling approach, model output type, and situational awareness from the perspective of an AV.
1 July 2024 in IEEE Transactions on Field Robotics
The development of a fully autonomous driving vehicle (AV) requires various traffic situations to be handled efficiently. One of the most common driving manoeuvres which an AV experiences in daily traffic is giving way (yielding) to other traffic participants. In this paper, we propose a simple yet efficient method of yielding that doesn’t query yielding areas of interests from map API making it hassle free to use without having to rely on digitized yielding areas. We incorporated our method into one of the well-known open-source autonomy stacks called Autoware. The proposed method makes use of high-definition (HD) map elements including lanes and stoplines for filtering vehicles which participate in yielding decision making. Our method estimates future collisions of filtered vehicles of interest with AV’s planned trajectory and outputs a binary yielding decision for ego vehicle. Our method covers different yielding areas including a roundabout and an unprotected turn. We tested and evaluated the decision making of our method on various simulated scenarios and afterwards successful real-world tests were conducted using an in-house AV. An in-depth analysis of our approach shows that the proposed yielding solution works reasonably well i.e. 87% successful yielding area navigation ratio on real data.
3 July 2025 in IEEE Robotics and Automation Letters
Predicting future trajectories of surrounding vehicles heavily relies on what contextual information is given to a motion prediction model. The context itself can be static (lanes, regulatory elements, etc) or dynamic (traffic participants). This letter presents a lane graph-based motion prediction model that first predicts graph-based goal proposals and later fuses them with cross attention over multiple contextual elements. We follow the famous encoder-interactor-decoder architecture where the encoder encodes scene context using lightweight Gated Recurrent Units, the interactor applies cross-context attention over encoded scene features and graph goal proposals, and the decoder regresses multimodal trajectories via Laplacian Mixture Density Network from the aggregated encodings. Using cross-attention over graph-based goal proposals gives robust trajectory estimates since the model learns to attend to future goal-relevant scene elements for the intended agent. We evaluate our work on nuScenes motion prediction dataset, achieving state-of-the-art results.