Detection of Generic Human-Object Interactions in Video Streams

Abstract

The detection of human-object interactions is a key component in many applications, examples include activity recognition, human intention understanding or the prediction of human movements. In this paper, we propose a novel framework to detect such interactions in RGB-D video streams based on spatio-temporal and pose information. Our system first detects possible human-object interactions using position and pose data of humans and objects. To counter false positive and false negative detections, we calculate the likelihood that such an interaction really occurs by tracking it over subsequent frames. Previous work mainly focused on the detection of specific activities with interacted objects in short prerecorded video clips. In contrast to that, our framework is able to find arbitrary interactions with 510 different objects exploiting the detection capabilities of R-CNNs as well as the Open Image dataset and can be used on online video streams. Our experimental evaluation demonstrates the robustness of the approach on various published videos recorded in indoor environments. The system achieves precision and recall rates of 0.82 on this dataset. Furthermore, we also show that our system can be used for online human motion prediction in robotic applications.