PEC: Human-robot collaborative dataset for behavior recognition in pathology examination scenes

In recent years, the number of iterations required to complete innovative research experiments has been increasing with the advancement of biomedical technology. Existing data-driven methods based on industrial datasets are difficult to apply to pathological experiments in biomedicine directly. We propose a pathological experiment collaborative dataset (PEC dataset) in our research to integrate robotic technology into pathological experiments and unleash its high efficiency. This dataset can provide a large amount of data for training the robot’s perception and collaboration capabilities. Based on three standard pathological experiments, we conducted a total of 180 experimental operations and collected RGB images and depth images as the foundation of the dataset. Correspondingly, we created 2D and 3D label files to record human action intentions and motion trajectories. To verify the adaptability of the proposed dataset, two general action and motion predicting methods were used to do some baseline testing tasks. Based on these two baseline methods, we evaluated the dataset and obtained an accuracy of over 80% in the continuous action prediction experiment. The average prediction error was kept within 50mm in the motion trajectory experiment. The experimental results demonstrate that, with training on the PEC dataset, robots can effectively acquire human motion intentions in pathological experimental environments, facilitating improved collaboration.