Gaze estimation provides insight into a person's intent and engagement level, which is helpful in collaborative human-robot applications. With significant advancements in deep learning architectures, appearance-based gaze estimation has gained much attention. Appearance-based methods have shown significant improvement in gaze accuracy and, unlike traditional approaches, they function well in environments where there are no constraints. We present another convolution-based gaze estimation approach to further reduce the angular error. For estimating gaze under extreme conditions such as head variations and distances, full-face images have been shown to be efficient, so we rely on full-face and pay more attention to necessary features. With the proposed architecture, we achieve an accuracy of 3.75° on the MPIIFaceGaze dataset and 3.96° on the ETH-XGaze open-source dataset. In addition, we test eye gaze tracking in real-time robotic applications, such as attention detection, and pick-and-place.

In this paper, we consider the problem of supporting modern financial exchange services on the cloud premises. Important exchange services rely on predictable, equal latency from the servers to the participants for fair competition. Existing cloud networks, however, are unable to offer such property, as they were not originally designed for this purpose. We attempt to tackle the problem of unfairness that stems from the lack of determinism in cloud networks. We argue that predictable or bounded latency is not necessary to achieve fairness. Inspired by the use of logical clocks in distributed systems, we propose a new approach that instead corrects for differences in latency to the participants for fairness. We evaluate our approach in simulation and show that it is feasible to achieve fairness under highly variable network latency. Our approach is deployable in contemporary cloud environments; it avoids limitations of state-of-the-art and outperforms it.

Motion prediction and action recognition play an influential role in the enhancement of interactions between humans and robots. We aim to predict motions and recognize actions for an interaction-based supermarket assistance scenario. Skeleton-based prediction of human motion and action recognition methods gained a lot of attention with the help of recurrent neural networks, convolutional neural networks, and graph convolutions. For recognition of actions, most of the proposed architectures rely on the predefined structure of the skeleton. In this work, we introduce a new small-scale dataset with actions that are possible in a supermarket interaction scenario. we propose two different self-attention-based models for recognition of actions for learning long-range correlations that do not rely on a predefined skeleton structure. We evaluate the models with extensive experiments containing specific input feature encodings that enhances the motion or trajectory features for accurate prediction and recognition of actions. We validate the effectiveness of the models on the actions in supermarket dataset and a standard benchmark dataset for action recognition known as the NTU RGB+D dataset.