Egocentric-Vision based Hand Posture Control System for Reconnaissance Robots

To facilitate full-loaded commandos to control reconnaissance robots, in this paper, we propose a wearable hand posture control system based on egocentric-vision by imitating the sign language interaction way among commandos. Considering the characteristics of the egocentric-vision on the battlefield, such as complicated backgrounds, large ego-motions and extreme transitions in lighting, a new hand detector based on Binary Edge HOG Block (BEHB) features is proposed to extract articulated postures from the egocentric-vision. Different from many other methods that use skin color cues, our proposed hand detector adopts contour cues and part-based voting idea. This means that our algorithm can be used on the battlefield even in dark environment, because infrared cameras can be used to get contour images rather than skin color images. The experiment result shows that the proposed hand detector can get a better posture detection result on the NUS hand posture dataset II. To improve hand recognition accuracy, a deep ensemble hybrid classifier is proposed by combing hybrid CNN-SVM classifier and ensemble technique. Compared with other state-of-art algorithms, the proposed classifier yields a recognition accuracy of 97.72 % on the NUS hand posture dataset II. At last, to reduce misjudgments during consecutive posture switches, a vote filter is proposed and applied to the sequence of the recognition results. The scout experiment shows that our wearable hand posture control system is more suitable than traditional hand-held controllers for full-loaded commandos to control reconnaissance robots.