BehaveAI: a framework for rapidly detecting and classifying objects and behaviour from motion

Here we introduce BehaveAI, a biologically inspired video analysis framework that integrates static and motion information through a novel colour-from-motion encoding strategy. This method translates object movement - direction, speed, and acceleration - into colour gradients, enabling both human annotators and pre-trained convolutional neural networks (CNNs) to infer motion patterns while retaining high-resolution spatial detail. Using a range of case studies, we demonstrate how the increased salience of motion information allows for the robust detection of objects that are challenging or impossible to identify reliably from static frames alone, particularly in complex natural scenes. We further demonstrate the reliable classification of different behaviours in animals and single-celled organisms. Additionally, the framework supports flexible hierarchical model structures that can separate the tasks of detection and classification for optimal efficiency, and provide individual tracking data that specifies what is present where and what it is doing in each frame. The framework makes use of the latest deep learning architecture (YOLO), combined with a semi-supervised annotation workflow. Together with salient motion information, these features can dramatically reduce the effort required for dataset annotation such that reliable models can often be made within an hour. Moreover, smaller annotation datasets mean that model training can be achieved on conventional computers without dedicated hardware, thereby improving accessibility. The motion encoding approach is also computationally lightweight, and can run in real-time on low-end edge devices such as a Raspberry Pi. We release the framework as a free, open source, and user-friendly package.