Extraction and classification of moving objects in robot applications using GMM-based background subtraction and SVMs

Object detection and classification is a common problem in industrial robot applications with the support of a computer vision system. The computer vision system detects and recognizes objects in the workplace, and extracts the features in real-time to provide feedback information for robot control. Support vector machine (SVM) is a supervised machine learning commonly used in classification problems as it produces notable correctness with less computation power. However, utilizing SVM to recognize an object in an image of the whole workspace is very time-consuming since we must divide the entire image into many small parts, extract features to build SVM for recognizing the object in each part. In this paper, to apply SVMs for a custom robot system with limited hardware performance, a background subtraction based on Gauss Mixture Model (GMM) is present to localize the exact position of the moving objects on a belt conveyor for a robot application. Then, the regions of objects are cropped and features in these regions are extracted for building an SVM for classifying objects based on their shape. Because only regions that contain objects are processed for SVM training and predicting, the proposed method can be applied in a robot system with a limited computational ability of a low-cost computer. The performance of GMM-based background subtraction is evaluated with two conditions: illumination changes and change in object movement velocity. The SVM is built to classify objects into eighths classes with different shapes. The experiment results reveal that SVM gives a high classification accuracy with the lowest accuracy of 98%.