ActionAI is an open-source Python toolkit for real-time human activity recognition that detects what people are doing by tracking body keypoints over time. It combines pose estimation with temporal classification so activities are localized both spatially and temporally in a video stream โ and it is designed to run on small edge devices rather than servers. The project is hosted on GitHub, where it has earned roughly 810 stars, with its most recent updates in 2024.
Recognizing actions in video usually requires heavyweight models and cloud GPUs. ActionAI takes a lighter route: it extracts skeletal keypoints from each frame, tracks multiple people through the scene, and feeds the keypoint sequences into sequence models such as LSTMs to classify activities as they happen. Because the pipeline works on compact pose representations instead of raw pixels, it can run in real time on hardware like the NVIDIA Jetson Nano, Raspberry Pi and Google Coral EdgeTPU, which makes it practical for privacy-sensitive, on-device monitoring. The codebase builds on TensorFlow 2 and scikit-learn. Its authors demonstrated the system in smart-fitness and monitoring scenarios, showing that useful activity recognition does not have to wait for a datacenter GPU.
ActionAI is free and open source under the GPL-3.0 license. There is no commercial edition or hosted offering โ you run it on your own edge hardware.
ActionAI fits makers, researchers and embedded-AI engineers who need activity recognition โ for safety monitoring, smart spaces or interactive installations โ running locally on inexpensive edge devices.