Vision based hand gesture recognition takes use of camera to capture image sequence, through image preprocessing and feature analysis to recognize and classify hand gestures, the extraction of features has direct relationship with the purpose and method of recognition. The commonly used hand gesture feature extraction techniques such as HOG (Histogram of Gradient) and image subspace projection, they are not only liable to fail when the background is cluttered but also need the training of a large number of samples, on the basis of former experiences, a new method based on density convex defect detection and density distribution is proposed in this paper, convex defect features and gesture distribution features are integrated to describe the characteristics of hand gesture including fingertips, contour length and area, density distribution and relative distance between fingers. The feasibility of the features proposed are proved by choosing 5 kinds of number gestures with total 500 images as test images for hand gesture recognition experiences, results shows that proposed approach is invariant to rotation, scale and translation, and it is more simple and precise than other similar methods.