Background

Multi-person action anticipation is pivotal for applications like autonomous driving, surveillance, and human-computer interaction. Traditional methods rely on object detection for classifying entities in a scene and use these classifications to construct graphs, with RNNs predicting future actions. These approaches, however, face challenges such as computational complexity, reliance on accurate object detection, and handling dynamic scenes. To address these challenges, this thesis explores the development of methodologies that predict future actions directly from observed data, without traditional graph construction. This aims to reduce computational demands and minimize errors associated with object detection, paving the way for more efficient and accurate predictions in dynamic multi-person environments.

Tasks

• Validation and benchmarking of State-of-the-Art results.
• Investigation of advanced spatial-temporal modeling
• Development of graph-free multi-person future state forecasting.

References

[1] Sun, C., Shrivastava, A., Vondrick, C., Sukthankar, R., Murphy, K., & Schmid, C. (2019). Relational action forecasting. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (pp. 273-283).

[2] Li, Y., Wang, P., & Chan, C. Y. (2021). RESTEP into the future: relational spatio-temporal learning for multi-person action forecasting. IEEE Transactions on Multimedia.

Conditions

• Subject: computer science, mathematics, electrical engineering, applied physics, mechatronics with good programming skills
• Willingness to familiarize yourself with new topics and enjoy bringing in your own ideas
• Good English or German speaking and writing skills, ability to work independently, and strong analytical skills
• Good understanding of deep learning basics and experience with DL projects

We Offer

• Intensive support and a pleasant working atmosphere in a creative team of motivated scientists
• Possibility of a subsequent job as a research assistant in order to further deepen the knowledge acquired
• Development of joint publications

Contact

Please send an email to zeyun.zhong@kit.edu with your transcript of records.