Human activity recognition (HAR) is a promising research issue in ubiquitous and wearable computing. However, there are some problems existing in traditional methods: 1) They treat HAR as a single label classification task, and ignore the information from other related tasks, which is helpful for the original task. 2) They need to predesign features artificially, which are heuristic and not tightly related to HAR task. To address these problems, we propose AROMA (human activity recognition using deep multi-task learning). Human activities can be divided into simple and complex activities. They are closely linked. Simple and complex activity recognitions are two related tasks in AROMA. For simple activity recognition task, AROMA utilizes a convolutional neural network (CNN) to extract deep features, which are task dependent and non-handcrafted. For complex activity recognition task, AROMA applies a long short-term memory (LSTM) network to learn the temporal context of activity data. In addition, there is a shared structure between the two tasks, and the object functions of these two tasks are optimized jointly. We evaluate AROMA on two public datasets, and the experimental results show that AROMA is able to yield a competitive performance in both simple and complex activity recognitions.

No abstract

In this paper, we give an experiment to investigate the change of physiological signals during playing body-controlled games. The physiological signals, including pulse rate, skin temperature, saturation of peripheral oxygen (SpO2), and galvanic skin response (GSR), of eleven healthy participants were recorded while playing body-controlled games. Based on the results of the experiment, we propose a discriminant model to predict the fatigue state of players. Our model can identify non-fatigue with 78.90% accuracy and fatigue with 82.76% accuracy. This model can be used with biofeedback hardware to continuously predict players' fatigue state and to improve the adaptation design of body-controlled games.

No abstract

Geo-tagged photo based tourist attraction recommendation can discover users’ travel preferences from their taken photos, so as to recommend suitable tourist attractions to them. However, existing visual content based methods cannot fully exploit the user and tourist attraction information of photos to extract visual features, and do not differentiate the significance of different photos. In this paper, we propose multi-level visual similarity based personalized tourist attraction recommendation using geo-tagged photos (MEAL). MEAL utilizes the visual contents of photos and interaction behavior data to obtain the final embeddings of users and tourist attractions, which are then used to predict the visit probabilities. Specifically, by crossing the user and tourist attraction information of photos, we define four visual similarity levels and introduce a corresponding quintuplet loss to embed the visual contents of photos. In addition, to capture the significance of different photos, we exploit the self-attention mechanism to obtain the visual representations of users and tourist attractions. We conducted experiments on two datasets crawled from Flickr, and the experimental results proved the advantage of this method.