Road user behavior prediction
| Title | Road user behavior prediction |
|---|---|
| Summary | Road user behavior recognition and manipulation using deep learning |
| Keywords | Deep learning, Intention recognition, behavior prediction |
| TimeFrame | HT2022-VT2023 |
| References | |
| Prerequisites | Machine Learning |
| Author | |
| Supervisor | Cristofer Englund, Björn Åstrand, Fernando Alonso-Fernandez |
| Level | Master |
| Status | Open |
Automated vehicles should be programmed to follow traffic rules, however, how do they cope in real traffic where rules may be ambiguous or a driver signs to other road-users to communicate. We believe that human drivers interact and communicate with their behavior i.e. speed, lane position and acceleration and deceleration patterns. How can this behavior be explained to a human, and how can we bild automated vehicles that can interpret the behavior of humans?
We have collected trajectory data from two different scenarios. One in traffic and one at a construction yard. In traffic the data is from an urban road that has a narrowing, so that vehicles cannot meet. One of the vehicles needs to stop to let the other pass. At the construction yard, trucks come in to get loaded or unloaded by forklifts. In addition, persons are walking in the yard and needs to interact with the moving vehicles. The trajectories consist of samples of time, x,-coordinates, speed, heading and road user type. Previous research concerned looking at only one vehicle, not interacting with anyone else (https://doi.org/10.1504/IJVD.2020.115056 PDF). Now we want to study the mutual relationship between the vehicles. If one vehicle is accelerating, how does that affect the other? If one vehicle moves to the left, will the other yield?
Different approaches to the problem exist. Collision-Free LSTM is one approach (https://doi.org/10.1007/978-3-319-73603-7_9). Another LSTM approach is Social LTSM: Human Trajectory Prediction in Crowded Spaces (https://ieeexplore.ieee.org/stamp/stamp.jsp?tp=&arnumber=7780479) Maneuver-Aware Pooling for Vehicle Trajectory Prediction: https://arxiv.org/pdf/2104.14079.pdf
Reinforcement learning is another method to take on the identification of behavior and movements. In (https://doi.org/10.3389/frai.2021.550030) a review of explainable AI and Reinforcement Learning is presented that can give inspiration to the given problem.
Another field of research that may be of interest for this domain is transformers (https://openaccess.thecvf.com/content/CVPR2021/html/Liu_Multimodal_Motion_Prediction_With_Stacked_Transformers_CVPR_2021_paper.html)
Does this sound interesting?
Contact us for further information.
