Gesture Recognition in Complex Background Based on Fusion of Hog and Faster R-CNN

Gesture recognition in simple backgrounds performs well, but it is still challenging in complex background. Because the gesture is indistinguishable from the surrounding background, gesture recognition becomes difficult. To reduce the impact of complex backgrounds on gesture recognition, we decided to adopt target detection to extract gestures area following by gesture recognition. So, we propose a gesture recognition network based on fusion of Hog and Faster R-CNN. By introducing HOG-SVM module, we fine-tune the loss value of the network to enhance the accuracy of gestures. Then, SE (Squeeze and Excitation) spatial attention mechanism is introduced to improve the accuracy of prediction bounding box. The results of the experiment show that the accuracy can reach 95.8% on PostHand dataset, which is 2.4% higher than Faster R-CNN.