Human and Robot Localization Using Histogram of Oriented Gradients (HOG) Feature for an Active Information Display in Intelligent Space

Intelligent Space is a well configured environment to provide various services for human and robots. This paper describes a vision-based localization of human and robot for an active information display service in Intelligent Space. For this, we configure a multi-camera network as a method to estimate 3D position of human and robot. In this space, human and robot are detected by using the Histogram of Oriented Gradients (HOG) feature method since it shows good performance for detecting human and robot. To train the HOG feature set, we have adopted an Adaptive Boosting (AdaBoost) learning algorithm, which chooses the best boosted classifiers of all generated classifiers. For the purpose of detecting human and a robot, we have used a linear Support Vector Machine (SVM), whereas Radial Basis Function SVM is used as a classifier for high accuracy. By using the multiple view geometry, the system estimates their 3D position from the positions of human and robot on image coordinates. To test the performance of the system, the robot or human is asked to track a rectangular trajectory and measure the error between the command and executed trajectories. Result shows that the average localization error is about 5.2 cm for 200 cm length of one side of the square, which is fairly accurate and good for a practical application of an active information display.