University of Tehran Mathematics Students Shine at SHL 2026 International Challenge

According to the Public Relations Office of the University of Tehran, the team from the Faculty of Mathematics, Statistics, and Computer Science of the University of Tehran, supervised by Dr. Bagher Babaali, Associate Professor of the Faculty, won first place in the Sussex-Huawei Locomotion Challenge 2026 (SHL) and became champion of this scientific challenge for the second consecutive year.

The team members are Mohammad Mahdi Mohebbizadeh, Ali Majlesi, Amir Mohammad Basouti, Iraj Gharai, and Mohammad Mahdi Azizi, graduate students in artificial intelligence at the University of Tehran.

The Sussex-Huawei Locomotion Challenge is one of the well-known international competitions in the field of Human Activity Recognition (HAR), organized in collaboration with the University of Sussex, UK, and supported by Huawei. This annual competition is held as a specialized workshop alongside the UbiComp/ISWC international conference. The 2026 edition is being held in Shanghai, China, from 19 to 23 Mehr (11–15 October 2026).

Accurate recognition of human activities using data from wearable sensors and smartphones is an important area of artificial intelligence research, with wide-ranging applications in health monitoring, analysis of movement patterns, location- and context-aware smart services, and transportation and personal mobility systems. Relying on a large-scale dataset of smartphone sensor data collected under real-life conditions, the competition provides a platform for evaluating and comparing modern AI methods in this field. Further information about the competition is available on its official website at http://www.shl-dataset.org/challenge-2026.

The 2026 research challenge focused on developing activity recognition methods based on Time-Series Foundation Models. One of the main constraints of the competition was keeping the foundation models frozen throughout all stages of system development, meaning that participants were not allowed to retrain or modify the core parameters of these models.

To address this constraint, the University of Tehran team designed and implemented an approach based on combining multiple foundation models and intelligently fusing their information. The method used five time-series foundation models: Mantis, NuTime, TiRex, TSPulse, and UniShape. By designing lightweight, dedicated classification layers, team members enabled effective use of the representations extracted by these frozen models.

The designed system included preprocessing of raw IMU data, non-overlapping 5-second windowing, 0.3–20 Hz band-pass filtering, and normalization; extraction of representations from five frozen foundation models; and finally classification and evaluation using dedicated heads and fusion strategies. Two main information fusion strategies, Early Fusion and Late Fusion, were extensively evaluated.

The results led to the selection of a Global Linear Fusion strategy, an approach that combines the outputs of different models to leverage their complementary capabilities and helps achieve competitive performance under diverse sensor data conditions. Technical details of the method and the team’s research findings are presented in the scientific paper associated with the competition.