| Summary: | In this study, we developed a framework to localize human lying poses by a camera positioned above. Our framework is motivated by the fact that detecting lying poses is fundamentally more difficult than detecting pedestrians or localizing nondeformable objects such as cars, roads, and buildings due to the large number of poses, orientations, and scales that a human lying on the ground can take. An important problem with lying pose detection is the training dataset, which hardly accounts for each possible body configuration. As a solution, we propose a geometric expansion procedure that uses a virtual camera to increase the number of training images. We also use a Gibbs sampler to generate more training samples in the feature space on which the system can train its model. Once the training is completed, detection is performed on a multiscale and multirotational space. Because our framework accommodates a variety of object detection systems, we report the results for the Faster R-CNN, FPN, and RefineDet models. The results show that using automatic dataset expansion models systematically improves the results.
|
|---|
