Gesture Recognition Based on Hexagonal Structure Histograms of Oriented Gradients
International Journal of Signal Processing, Image Processing and Pattern Recognition
Feature extraction methods of image directly affect the feature recognition results based on computer vision-based gesture recognition. This paper proposed a gesture feature extraction method of hexagonal structure, which is based on histograms of oriented gradients. This paper transforms the quadrilateral structure into hexagonal structure of the image and defines hexagonal block structure. Using different structures, different sizes of hexagonal blocks extract gesture feature and recognize on
... different sizes image. Experimental results show that hexagonal structure block is more appropriate than conventional histograms of oriented gradients structure block, the hexagonal structure histograms of oriented gradients feature extraction method is more efficient than the square and circular structure histograms of oriented gradients feature, using gentle_adaboost classifier achieved higher recognition rate. Chen and Tseng  proposed a multi-angle hand-gesture recognition method in their article. They used three webcams set at, front, right, left directions of hand to capture gestures. Then three SVM (Support Vector Machine) classifiers trained respectively. After the training process, one voting and two plans of fusion fused the constructed classifiers. The recognition rate of hand gestures more than 93%, including different angles, sizes, and different skin colors. However, there were only three hand gestures in their research. Huang and Hu  applied the method of gabor filter, PCA and SVM to recognize many simple hand gestures in complex background. They first extracted the hands from a sequence of video images using the skin color information. Then, they coped with these images of hand gestures using gabor filter, and they used PCA method to reduce the dimension of the data space and used SVM classifier to recognize the hand gestures. The recognition rate of 95.2% can be achieved. The experiment results show that the processing time is 0.2 second for every frame, which did not achieved the requirement of a real-time system. Huang and Hu  estimate the orientation of the hand gestures using the gabor filter responses. The estimated angle is used to correct the hand pose into an upright orientation. Amin and Yan  proposed a system that is able to recognize ASL (American Sign Language alphabets) from hand gesture with average 93.23%. They used PCA and gabor filters to accomplish the task, out of the top 20 principal components the best combination of principal components is determined by finding the best fuzzy cluster for the corresponding PCs of the training data, their experiment demonstrate the best result obtained from the combination of the fourth to seventh principle components. However, their recognition rate of similar alphabets is relatively low. Jayashree R  uses fixed position low-cost web camera with 10 mega pixel resolution mounted on the top of monitor of computer which captures snapshot using RGB color space from fixed distance. This gesture recognition system can reliably recognize single-hand gestures in real time and can achieve a 90.19% recognition rate in complex background with a "minimum-possible constraints" approach. Li  proposed a speed hand gesture recognition system using the kinect sensor. Based on the HOG features and adaboost training algorithm, the experiment demonstrated the detection results is great, but for the situations like hands covered in front of body or objects kind of similar to hands, there is still some high missing and false rate. The application of hand gesture recognition for real-life is very challenging because of the requirements on its robustness, accuracy and efficiency. Padam Priyal  has presented a gesture recognition system using geometry based normalizations and Krawtchouk moment features for classifying static hand gestures. The proposed system is robust to similarity transformations and projective variations. Cao  and Zhang  using image for depth extraction and recognition of shape matching method in hand gesture recognition has obtained the certain effect, but susceptible to the influence of occlusion. The literatures   combined with histogram of oriented gradient and local gradient direction for recognizing gesture, and achieved good results. The literature  identify the hotspot function of the workload on an embedded system that motivates acceleration and present the detailed design of a hardware accelerator for histograms of oriented gradients descriptor extraction. The literature  incorporated fuzzy concept to HOG aiming to achieve a good recognition rate with a low feature vector dimension. Experimental results have demonstrated that HFOG outperforms the original HOG with a lower dimensional vector. But HOG structure affected the classification performance. In our article, we had proposed an algorithm of hand-gesture recognition based on H-HOG descriptor and Gentle_adaboost classifier, effectively addressing the static hand-gesture recognition problem. At first, the gesture images were transforming into hexagon images. Then we obtained H-HOG features when hexagon block edge length and step length were taken different values. Further, we trained and tested H-HOG features. At last, we compared R-HOG, C-HOG and H-HOG features performance. Additionally, the recognition time fit to the requirement of a real-time system. Further, it is robust to nonlinear illumination and image blurring. At present, 12 hand gestures had been used in our experiment. Some hand-gesture images were very similar; therefore, the classification error rate of our method was relatively high. Further research needed to research new feature descriptor.