A hybrid active vision system for real time application running object recognition

Abstract

Computer vision is developing fast and be applied to robotics. Active vision system has been researched recently to address problems with saccade movement and real-time realization. The paper proposed the use of a hybrid system consisting of an object recognition model built with MobileNet SSD and an object tracker running Median Flow. This system has proven to work well in real-time application, with an average 15 frames per second and angular speed of 75 degree per second. The system remains stable and robust throughout all experiments. The final design is implemented to control movement of a transhumeral prosthesis through determining object of interest, calibrating wrist according to object position and choosing grip patterns with respective to object characteristics.