Research For Human Detection Based On Depth Image

Human detection is the foundation of many applications such as human tracking, action recognition, human compute interaction and augmented reality. With the release of several consumer depth sensors like Kinect and Realsense, it is feasible to obtain the depth data. Moreover, those depth images are insensitive to illumination variations. As a result, research on detection algorithms on depth images has attracted ever increasing attentions.The purpose of human detection algorithm is to detect human as good as possible. This can be achieved with two ways. First way is focusing on designing more discriminative features, another way is to fusion several features. In this paper, we will research human detection based on depth images from these two sides.Firstly, the research of the feature extraction on depth images has been highlighted in this paper. We improved Local Directional Patterns(LDP) feature used in RGB images at first, we changed the unsigned LDP value to signed value with the fact that the sign of the LDP value denote depth difference direction. Then in consideration of that the existing methods usually focused on encoding the local feature distribution without any human body spatial structure information, we proposed an encoding framework that suitable for human spatial structure characteristics. The experiments showed that the proposed method can achieve a higher detection rate than any other detection methods on depth images.On the other side, on the basis of the proposed detection method, we researched on feature fusion from the feature level fusion and the decision level fusion. In the feature level fusion, we used canonical correlation analysis method to analyze the correlation between two features and then formed them into a more discriminative feature. In the decision level fusion, we constructed a graph and used graph propagation to optimize it. Moreover, we proposed a framework to combine these two feature fusion methods to form a multi-level fusion method. The experiments showed that the proposed multi-level fusion method is better than any single feature method or any single level fusion method.Finally, we built a smart car following system based on the proposed detection method. With the detection result, the system can calculate the speed and orientation of the moving person and make the smart car to follow the motion. It showed the potential application value for the human detection based on depth images.