Advanced Computer Vision Project
Applying the Constellation of parts method to video
Constellation of parts is a compact and expressive means of modeling classes of objects. This model can be used to simultaneously segment and recognize/classify objects in images. The models can be “trained” by machine learning techniques and is invariant to scale and similar factors. For my project in Advanced Computer Vision at RIT, I will implement a constellation of parts algorithm to recognise objects in video.
Data collection
Videos of people walking from a still or moving camera. Outdoor scene with complex background. Variety of foreground busyness: no people (background only), one person, a few people, many people (moving in different directions and occluding each other). Variety of person behavior.
Implementation
Implement a constellation of parts model and algorithm to segment and recognize humans in video. Use training videos to develop the model (or build the model by hand if pressed for time). Implement the algorithm in OpenCV (C++). Measure results of segmentation/recognition subjectively against ideal results and objectively against the results of recent research (see references below). Evaluate the effectiveness and efficiency of the implementation on video.
Possible enhancements include the use of graph cuts to improve segmentation accuracy, etc.
References
Fergus, P. Perona, and A. Zisserman. “Object Class Recognition by Unsupervised Scale-Invariant Learning,” IEEE, 2003.