Jordi Pagès graduated in Computer Science in the University of Girona, Catalunya, Spain, in 2001. He received the Ph.D in Computer Vision and Robotics from the University of Girona and the University of Rennes, France, in 2005. He worked in Davantis Technologies in Barcelona developping intelligent video surveillance software. In 2009 he moved to Pal Robotics, where he is now a research scientist working in computer vision applied to humanoid robots.

My current position at Pal Robotics offers me the chance to work in humanoid robot perception. Many computer vision problems have to be considered such as:

• Object detection/recognition and 3D perception for robotic manipulation and obstacle avoidance
• People detection and tracking for making the robot aware of people around
• Visual slam, visual odometry and visual place recognition for robot navigation
• Gesture/action recognition for human-robot interaction

A huge number of surveillance cameras have been installed all over the world during the last decades. Nevertheless, keeping an eye on all these cameras is not practical if not impossible. Automating the detection of potential attacks, intrusions or suspicious actions can be achieved by using computer vision and artificial intelligence algorithms. In this project I worked with topics like:

• Background subtraction and motion detection
• Object tracking
• Object classification using heuristics and machine learning and data mining
• Event detection

In this work we face the problem of positioning a camera attached to the end-effector of a robotic manipulator so that it gets parallel to a planar object. This vision-based task is a classic application in robotics. Most part of visual servoing approaches solving the plane-to-plane positioning task are based on tracking some visual features from the object. We focus on the case of a markerless planar object where it is not possible to extract primitives like points, lines or regions.

The solution that we propose is to attach a structured light emitter to the camera so that it projects visual marks onto the object surface which can be used the control loop in order to fulfill the plane-to-plane positioning task. This work shows that structured light allows not only non-textured or markerless objects to be treated but also to optimize the control loop obtaining nice properties like decoupling degrees of freedom, stability and good camera trajectory.

This researh was made at the University of Girona along with the Cemagref of Rennes and the Lagadic team of the INRIA (France).

Projecting structured light patterns onto the environment is a largely used technique in computer vision and robotics. The main advantage is that the projected visual features are easily distinguished by a camera. In order to avoid ambiguities and reliably solving the correspondence problem the patterns can be coded. Coded structured light is a technique based on projecting a light pattern and imaging the illuminated scene from one or more points of view. Since the pattern is coded, correspondences between image points and points of the projected pattern can be easily found. During this project a comprehensive survey of the existing techniques was developed. We implemented a set of representative techniques and comparative results were presented. The benchmark framework used was shape acquisition from a single or multiple shots. Different types of patterns and coding strategies were compared according to their 3D reconstruction performance and their ability to provide correspondences.

In this work we also presented a new coloured pattern which is able to obtain correspondences with a unique shot. The pattern is generated by a new coding strategy based on De Bruijn sequences. The new pattern is compared to similar existing patterns by reconstructing the shape of different objects and analysing the results from a quantitative and a qualitative point of view. The results showed that the pattern is able to increase the number of correspondences in a single shot without loss of accuracy compared to other similar patterns.