Hossein Hashemi

Having a three-dimensional reconstruction of the human body is an important issue for machine vision and can be used extensively in human-computer interaction, virtual reality, video surveillance and video games. While 2D human body pose estimation has progressed very well in the last few years, considerable attention has recently been paid to 3D human pose estimation problems.

This project improves an algorithm that performs 3D motion capture (MoCap) with a 2D image or video as the only input to the process.

Recognizing human movement and activities from video sequences or images has been a challenging task for several years. It has many applications, including video surveillance systems, human-computer interaction and robotics for human behavior characterization. These applications require a system to understand how a person is moving and what they are doing.

In this work, I will propose a new method to analyze the human body joints in every frame of video for classifying actions and daily human activities, based on reservoir computing approaches to multivariate time series classification. The results of this research can be useful for complex body posture recognition or social action recognition.

(How do you find the kiss scenes in the film Titanic?)

Nowadays, human life is becoming more automated, with robots and artificial intelligence used in all aspects of industry. What is so interesting in modern life is creating comprehensive and effective interaction between robots and humans. Predicting human posture is a novel and interesting topic for pioneering and innovative industries such as autonomous cars, robot cameramen, video content understanding and human-robot interaction.

Assessing human behavior with artificial intelligence and deciding according to this assessment affects the results and reactions of robots and machines. However, there are a lot of methods and algorithms to control the reactions and responses of robots, but they are not effective enough to capture reality in an acceptable time and at high speed. They are also not capable of considering and investigating reality in three dimensions.

Expanding this knowledge and improving this technology so that machines react correctly and precisely will be the purpose of this project. The results can be useful for autonomous cars when they encounter a moving object: predicting the next behavior of that object, which can be a human, and making the correct decision. In the field of the robot cameraman, the system should be capable of predicting and following a moving object precisely.

Multi-person 3D visualization and localization in the street based on a single video is in progress.

Most commercial eye gaze tracking systems are based on the use of infrared light. However, such systems may not work outdoors or may have a very limited head box in which they can work. This project proposes a non-infrared approach to track one's eye gaze with an RGB-D camera (in our case, Kinect). The proposed method adopts a personalized 3D face model constructed offline.

To detect eye gaze, our system tracks the iris center and a set of 2D facial landmarks whose 3D locations are provided by the RGB-D camera. A simple one-time calibration procedure is used to obtain the parameters of the personalized eye gaze model.

We compare the performance of the proposed method against the 2D approach using only RGB input on the same images, and find that the use of depth information directly from Kinect achieves more accurate tracking. As expected, the results from the proposed method are not as accurate as those from infrared-based approaches. However, this method has the potential for practical use with upcoming better and cheaper depth cameras.