Detecting, refining, and tracking complex objects are important perception tasks for intelligent systems. The next acts and decisions of the intelligent systems are made based on the information provided by the perception modules. The shape, the degree of freedom, or the range of motion can determine the complexity of an object. Their complexities at different levels generate some challenges in the perception of these objects.
In this talk, I will focus on the perception algorithms to detect, refine, and track complex objects using two different representations, namely low dimensional and point-wise (non-parametric), by combining multiple features from a single or multiple sensors. The differences between two groups of algorithms which represent the complex objects in the low dimensional space and point-wise will be explored, compared and analyzed.
On the one side, a multi-modal human detection algorithm which utilizes data from multiple sensors is going to be explained. In this method, the human is represented in the low dimensional space. Multiple sensor architecture is constructed to obtain visual and geometric cues from different sensors and to increase the detection rate of the human classifier. This architecture supplies complementary and redundant information to enhance the confidence level of the detection, so this approach increases the robustness of the system.
On the other side, a modified graph cut based method to refine complex objects, e.g. trail, human hand and face, if a rough estimation of the object is given by a low dimensional shape tracker will be described. Going beyond single layer graph cut framework, a multi-layer graph structure to incorporate low and high level observations in a joint way for the refinement of low dimensional human shape will be outlined.
Mehmet Kemal Kocamaz is a PhD Candidate in the Department of Computer Science of University of Delaware. His research focuses on the perception systems of Autonomous Ground Vehicles, and complex object detection, tracking, and refinement. He received his BS degree from University of Southern California, and his Master degree from Rensselaer Polytechnic Institute in Computer Science in 2005 and 2007, respectively. He is a member of Warthog team which is the winner of 2009 and 2011 Autonomous Challenge of Intelligent Ground Vehicle Competition (IGVC) among ~60 robots from 6 countries. He worked on the perception modules of the robot. He is one of the inventors of Multi-modal Pedestrian Detection system which fuses 1D range scan and 2D image based information for Mitsubishi cars.