Thursday 06 March 2025
Researchers have made significant strides in developing a system that can detect when a person is holding onto an object, a crucial aspect of everyday interactions. The team’s approach relies on analyzing grasp quality metrics extracted from object and hand poses, demonstrating a high accuracy rate in their evaluation.
The system works by reconstructing the interaction scene using a simulator called GraspIt!, which allows for the evaluation of a human’s grasp of an object. To do this, the simulator uses data provided by the DexYCB dataset, which consists of annotated images and pose data of human-object interactions performed by 10 human subjects.
The team’s approach involves estimating hand shapes and poses using computer vision techniques, allowing them to extract grasp quality metrics from the reconstructed scene. These metrics are then compared to ground truth data, which indicates whether a person is successfully grasping an object or not.
In their evaluation, the system demonstrated a detection accuracy of 89.3%, with false positive rates ranging from 3% to 29%. The team also observed that higher grasp quality metric values were more common in frames where the human subject was successfully grasping an object. This suggests that these metrics could be used as indicators of successful grasping.
The system’s performance varied across different objects, with some objects being detected more accurately than others. For example, the object least accurately detected, a bowl, is relatively small and can be stably grasped on its sides. In contrast, the object most accurately detected, a pudding box, is small and has a stable grasp.
The team’s approach shows promise for real-world applications, particularly in robotic handovers where objects are transferred from one person to another. Traditional methods of detecting contact rely on physical interactions, such as force or contact sensing, which can be costly and may not always accurately detect contact.
In the future, the researchers plan to adapt their system for real-time operation using computer vision techniques for pose estimation. They also hope to integrate their system with a robot-to-human handover system, where object release is triggered by contact detection, and compare its performance with force-based contact detection methods.
Overall, this research demonstrates the potential of grasp quality metrics in detecting human-object interactions, and could have significant implications for robotic systems that rely on accurate detection of contact.
Cite this article: “Detecting Human-Object Interactions with High Accuracy Using Grasp Quality Metrics”, The Science Archive, 2025.
Grasp Quality, Object Detection, Computer Vision, Hand Pose Estimation, Simulator, Graspit!, Dexycb Dataset, Robotic Handovers, Contact Detection, Human-Object Interactions
