Human Tracking System For Mobile Robot Using Hierarchial Features Descriptor

As the maturity of robotics, more and more kinds of intelligent service robot gradually come into the human daily life, human tracking has gradually became the key technology of robot. To solve problems of scale variations, temporary occlusion and deformation of the target in the process of human tracking, a human tracking system based on hierarchical features descriptor of the target was proposed. In addition, an associating visual graphical user's interface was developed. The method was verified by experiments conducted on the Pioneer3 with cameras. The research work mainly included the following aspects:(1) The improved Cam-Shift algorithm based on adaptive bandwidth kernel functionThe traditional Cam-Shift with fixed bandwidth kernel function cannot effectively handle rotation and deformation of the target. Aiming to resolve these problems, an improved Cam-Shift algorithm based on adaptive bandwidth kernel function was proposed. This algorithm based on the traditional Cam-shift algorithm used the scale and dominant direction of the object to correct the bandwidth kernel function and tracking window, which can enhance the robustness of rotation and scale and improve the real-time performance of the algorithm by reducing the number of iterations.(2) The SIFT feature matching algorithm based on compressive sensingThe computational complexity of SIFT feature descriptor computing stage was quite expensive and the dimensionality of the feature vectors was relatively high. For speeding up the SIFT computation, the SIFT features matching algorithm based on compressive sensing was proposed. Firstly, the wavelet transform matrix was used to realize the sparse representation of the 128 dimensional feature vectors. And then,sparse random matrix was projected to reduce the dimension of the 128 dimension vectors, which can reduce the complexity of the algorithm and improve the real-time performance of system.(3) The SIFT features matching algorithm based on Contourlet transformationThe traditional SIFT feature descriptor was only about the color gradient in the neighborhood of key point and it discarded relativity of the key position in the whole picture structure corresponding that gave rise to a low matching robustness when tracking similar areas. Therefore, this paper put forward an innovative Contourlet-SIFT feature matching algorithm. The SIFT key points were first extracted to conduct Contourlet transformation on peripheral areas in order to calculate the mean and standard deviation of the decomposition coefficient in each direction. Thenthe vector of overall texture description is constructed and the Euclidean distance of this low-dimensional vector provides references for prioritizing the matching pairs.The introduction of Contourlet transformation enhanced the description of target while improving the accuracy of tracking.(4) The target description based on hierarchical feature descriptorIn order to enhance the description ability of the target, this paper proposed a hierarchical features description mechanism. The head and shoulders or color features as the target outer characteristics were used to determine the ROI. The CS-SIFT or the Contourlet-SIFT was selected to realize the precise description of the pedestrian target according to the local similarity. By combining the global and local features, the description ability of the target was further enhanced, while the robustness of the tracking system of the mobile robot in complex environment was improved. When the target scale or direction changed, the outer characteristics were modified according to the SIFT feature matching result. When the target body was short of occlusion or deformation, the online updating strategy of local feature template ensured the stability of the inner feature. Extend Kalman Filter(EKF) was used to predict the target motion state when the target was completely occluded,which solved under the condition of complete occlusion of the target.In conclusion, the human tracking algorithm described above was verified by experiments conducted on the Pioneer3-DX robot platform. Experimental results showed that the improved Cam-Shift algorithm based on adaptive bandwidth kernel function could handle inadequate representation of target's color distribution and invariable target model. The SIFT features matching algorithm based on compressive sensing effectively improved the real-time performance of the system. The SIFT features matching algorithm based on Contourlet transformation enhanced the ability of feature points to describe the target and improving the real-time performance of the algorithm. The hierarchical feature description of targets effectively combined the global and local features of the target, which enhanced the robustness of the tracking system and made the robot tracking process more stable and efficient.