Research On Hand Detection And Recognition System By Random Forest

With the development of intelligent identification technology and the improvement ofour living standard, the HCI (Human Computer Interaction) method based on artificialintelligence is closer and closer to our daily life. Just the same as natural language, handgestures could be used to convey information and instructions as sign languages. In recentyears, people have proposed a number of algorithms according to hand detection and gesturerecognition. However, most of these methods are not perfect because of the disadvantages oflow precision, poor stability, low efficiency, or high demand for hardware and so on. In thispaper, an intelligent hand gesture detection and recognition system used in HCI is proposed.Aiming at gesture detection and recognition, efforts in this article will mainly be made in thefollowing aspects.(1) Pair-patch comparison features are used to describe both the training samples anddetection samples. The value of a pair-patch comparison feature is the difference in pixelvalue between any two points of a sample image. This feature description method could wellformulate the distribution of pixels in the image. To improve the robustness of the proposedmethod, the value of each point used to calculate the feature is got by taking the average in the3-neighbourhood of the point. At last, the feature vectors, which are composited by thesepair-patch comparison features, are used to train a classifier for gesture recognition.(2) The random forest classifier is proposed to train the classifier of hand gestures in thispaper. In every nod of each tree in the forest, the pair-patch comparison features work as asplitting criterion to classify the sample set. In this method, we train many decision trees inthe forest classifier first, and then establish the random forest hand gesture classifieraccording to the result of each decision tree classifier. According to the structure of the treeclassifier, the system could get a high efficiency both in training phase and detection phase.What’s more, the precision is also very high, because the random forest is a kind of ascendingstrong classifier.(3) A skin detection method in YCbCr color space has been adopted as a preprocessing inthe detection phase to improve the efficiency. Besides, a frame difference method is used todetect the region with an obvious motion in the image sequence. The two methods in theabove context could remove most areas that have nothing to do with moving hand gestures, sothat the hand gesture searching region is limited in a small area. At last, we detect handgestures in this region by the random forest classifier got in the training phase. These two preprocessing methods could greatly improve the precision and speed of the system, so as torealize the real-time performance.Finally, this paper proposed a hand detection and gesture recognition system by randomforest. This system is based on three gestures containing palm, fist and finger. By analyzingthe experiment result, this system could real-time detect hand and recognize three differentgestures only need a normal web camera and related computer programs.