Object Tracking from Video Sequence (using color and texture information and RBF Networks)
Supervisor: Dr. MohammadReza Karami
Advisor: Dr. Yasser Baleghi
By: Alireza Asvadi
Computer vision is a field that combines image processing and machine learning methods in order to understand the scene. Tracking is a basic task for several applications of computer vision, e.g., motion-based recognition, automated surveillance, video indexing, human-computer interaction, traffic monitoring, vehicle navigation, and so on. In recent years, the research of the visual tracking has arrived at a climax. In the present thesis, an efficient method for object tracking is proposed using color and texture features and two Radial Basis Function Neural Networks. In the first frame operator selects the object region. A surrounding area is selected as a background automatically. In the next step, the pixel-based color (Hue, Saturation, and Value) and texture (Average, Variance and Range) features are extracted from the object and background regions. The first RBFNN classifier is trained and tested with these features. The first RBFNN is employed for separating object from its surrounding background in the first frame. Color and texture features from surrounding background are applied to develop an extended background model. The detected object and extended background features are then used to train the second RBFNN. The trained RBFNN is used as a scoring function that gives higher scores to the object pixels. The second trained RBFNN is employed to detect object in subsequent frames while meanshift procedure is used to track the object location. By introducing and measuring two criterions the object size and model are updated. The performance of the proposed tracker is tested with many video sequences. The proposed tracker is illustrated to be able to track object and successfully resolve the problems caused by the camera movement, partial occlusion and gradual change in object and background color and texture. Also the proposed method provides object region in every frame that can be used in higher level processing such as object and activity recognition.
Keywords: computer vision, object tracking, radial basis function neural networks, texture, mean shift.
Click [results] to download all video results in the zip format containing flv files
The proposed method is shown by solid red line. NCC is shown by dash-dot green, Mean shift by dashed pink.
The code runs on Windows XP with MATLAB R2009b.
The code is outdated now.