Visual Navigation (thesis topic)
Object Recognition (summer intern, Microsoft Research, Redmond, WA,1998)
Video Analysis (summer intern, C&C Research Lab, NEC Corp., San Jose, CA 2001)

I am always interested in developing vision systems to meet the challenges from real world applications, such as robotics and multimedia understanding.
 

Movie Clips on 2D visual servoing

The navigation capability is essential for anything called robot. I believe that to achieve human-level navigation capability, the robot has to understand the 3D structure of the environment, no shortcut. 

My thesis work shows that during a navigation task, it is possible to robustly recover the 3D structure with an omnidirectional camera even in presence of abrupt motion and occlusions, and in turn, the recovered 3D structure can profoundly enhance the robot navigation performance.

Most existing robot servoing algorithms use image feedback directly without reconstruction. Some algorithms use structure from motion techniques without addressing the robustness problem. We find the robustness issue is vital in uncontrolled environments, such as outdoors.

Explicit 3D reconstruction makes the controller design or planning straightforward. And that is very important for the stability and flexibility of robot navigation systems.

But the reconstruction itself is difficult. The problem of structure from motion has been largely studied under controlled lab environment. The equally difficult problem is to find correspondence or tracking, which is the pre-request for most reconstruction algorithms. I showed that by properly representing the uncertainty of both tracking and reconstruction, it is possible to robustly recover 3D structures during challenging navigation situations, such as abrupt robot motion and occlusions. This result can lead to a visual navigation system that is completely without human assistance.

Tracking: we developed a tracking module that can track various kinds of features.  See [1] or [2] (the journal version) for details about tracking with an omnidirectional camera.

Structure from motion: We developed a point based SFM algorithm for catadioptric omnidirectional camera [3]. Some recent work are toward patch-based SFM [4].

Robust tracking and SFM: The technical approach and some preliminary results can be found in [5].

3D reconstruction based visual navigation: ongoing work.

We collect the video by driving the robot around the campus and then use theotilite to measure the ground truth of the campus buildings. We compare the reconstruction of our algorithm with the ground truth.

[1] P. Chang and M. Hebert, Omnidirectional visual servoing for human-robot interaction, IROS'98
[2] P. Chang and M. Hebert, First result on visual servoing with an omnidirectional camera, Advanced Robotics, Vol. 14, No. 3, 2000. (Journal version of [1])
[3] P. Chang and M. Hebert, Omnidirectional structure from motion, proceedings of IEEE workshop on omnidirectional vision, Hilton Head, 2000 
[4] M. Hebert, R. MacLanchlan and P. Chang, Experiments with Driving Modes for Urban Robots, SPIE proceedings 1999.
[5] P. Chang and M. Hebert, Robust tracking and structure from motion through sampling based uncertainty representation, IEEE ICRA'02

In the summer of 1998, I worked with Dr. John Krumm on Easyling project in Microsoft Research. We developed a novel object recognition algorithm based on color cooccurrence histogram [1]. Three patents are filed and pending.

[6] P. Chang and J. Krumm, Object recognition with color cooccurrence histograms, CVPR' 99.

In the summer of 2001, I worked in the C&C Research Laboratories of NEC USA Corp. With Dr. Yihong Gong and other researchers, we prototyped a video analysis system which can detect highlights from baseball game videos.

[7] P. Chang, M. Han and Y. Gong, Highlight detection and classification of baseball game video with Hidden Markov Models, International Conference on Image Processing, 2002