Section: Highlights of the Year

Highlights of the Year

This year Stars has proposed new algorithms in the domains of perception for activity recognition and semantic activity recognition.

• A new Re-Identification algorithm which outperforms the State-of-the-art algorithms while being adapted to real-world applications (i.e. it does not require the use of heavy manual annotations which is typical of metric learning algorithms). The remaining challenge is to be able to distinguish people who have similar appearance.

• A new generic action recognition algorithm which outperforms the State-of-the-art algorithms. This algorithm uses new action descriptors that enable finer gesture classification. An open issue is to get a real-time implementation with good enough performance. An extension of this algorithm has been devised for RGB-D cameras, which has been demonstrated in a real-life application, where a robot has to recognize people taking their meal (e.g. eating, drinking).

• New generic tracking algorithms, which can optimize the on-line tuning of tracking parameters and can operate at different temporal scales to recover from lost tracklets. These tracking algorithms have been validated on real world videos lasting more than a week. The utilization of such sophisticated algorithms is still complex and requires some more researches for their deployment in a large variety of applications.

• New tools to help modeling human activities of daily living. These tools enable to evaluate and improve activity recognition algorithms on long videos depicting the performance of older people living in a nursing home in Nice. The utilization of these tools by clinicians and medical doctors is an ongoing task.

• A new algorithm to recognize human activities, that can benefit from the fusion of events coming from camera networks and heterogeneous sensors.

• A new algorithm to discover human activities of daily living by processing in an unsupervised manner a large collection of videos. The generation of the event models does not require the use of heavy manual annotations which is typical of supervised activity recognition algorithms. However this algorithm still need to have well tracked people to be able to understand their behaviors with sufficient precision.