Wednesday 09 April 2025
The quest for a more accurate way to recognize human actions has been ongoing for decades, with researchers pouring over vast amounts of video data in search of the perfect solution. A recent challenge aimed at tackling this problem has seen some impressive results, with a team from L3i Laboratory, La Rochelle University, France, emerging as one of the top performers.
The Elderly Action Recognition (EAR) Challenge, part of the Computer Vision for Small Workshop at WACV 2025, focused on recognizing Activities of Daily Living (ADLs) performed by the elderly. The competition’s dataset consisted of six action categories – locomotion, manipulation, hygiene, eating, communication, and leisure – which present a unique set of challenges due to their real-world scenarios.
To tackle these challenges, the researchers employed a state-of-the-art video action recognition model, fine-tuned using transfer learning on elderly-specific datasets. This approach allowed them to enhance adaptability and improve generalization, making it more effective in classifying elderly activities.
The team’s solution was built around the Temporal Shift Module (TSM), which has been shown to excel in action recognition tasks. By leveraging this module with a resnext50 32x4d backbone, they were able to achieve impressive results on the public leaderboard.
One of the key aspects of their approach was the use of two different configurations for training and testing. The first configuration included the Toyota Smarthome dataset, along with RGB videos from the ETRI-Activity3D and ETRI-Activity3D-LivingLab datasets. This combination provided a solid foundation for training the model.
The second configuration used the full RGB videos from both datasets, which allowed them to further improve their results. By applying targeted pre-processing techniques and carefully curating the training data from multiple publicly available sources, they were able to mitigate dataset bias and achieve better generalization.
The results speak for themselves – the team’s solution currently holds an accuracy of 0.81455 on the public leaderboard, with a significant improvement in performance seen after further training. Their approach outperformed many other top-performing teams, including CUHK, RoboVision, VisionLab, and CVMI.
This achievement has important implications for elderly care, health monitoring, and assistive technologies. By developing accurate models for recognizing ADLs, researchers can create more effective systems that can better support the needs of older adults.
Cite this article: “Elderly Action Recognition in the Wild: A Deep Learning Approach to Improve Healthcare Outcomes”, The Science Archive, 2025.
Elderly Action Recognition, Computer Vision, Activity Recognition, Adls, Video Analysis, Machine Learning, Transfer Learning, Temporal Shift Module, Resnext50, Toyota Smarthome
Reference: Anh-Kiet Duong, “Elderly Activity Recognition in the Wild: Results from the EAR Challenge” (2025).
