Case-Study: Action Recognition


Our client had a security-related problem that required monitoring humans taking suspicious action within the confines of an area, using cost-effective CCTV installations.


With our initial approach, we attempted to design a solution more heavily relying on deep learning, but it turned out mid-project that the expected quantity and quality of training data could not be supplied. Our team managed to respond to the situation with the development of a hybrid model, where modelling of the human body could be done with transfer learning, whilst the suspicious actions could be identified in a rules-based way.

The resulting solution was developed in a way that allowed for GPU-backed processing of a PyTorch model and vectorized NumPy code for the rules-based parts.


Proof-of-concept delivery was deemed successful, considering the limitations of the available training data. The system in its current shape still requires human supervision, but successful commercial deployment could pave the way for the acquisition of an improved training set, too.

Image by Freepik
Image by Freepik