Research On Gait Object Extraction And Recognition

Gait recognition is the method of identifying individuals by their walk manners and gesture using computer technologies. Gait as a type of biometrical features have many advantages such as remote data acquisition, carrying out test without the awareness and using low resolution gait images, gait became the most potential biometrical features in remote recognition and became one the research focus. We discuss three important aspects: preprocessing of gait image, feature extraction and classification. We emphasis on silhouette segmentation algorithm and silhouette description using Tangent Vector Descriptor . On the base of previous studies, we have finished some research work as following:(1) Silhouette segmentation from gait video is of great importance in the field of gait recognition. The paper aims at putting forward a new kind of algorithm—gait object segmentation algorithm to get good accuracy and have fewer small holes and gaps without any shadow as well. With the help of using the vector generated by one pixel and its neighborhood ones as the representatives of the pixel, the similarities of pixels can be indicated by each two vectors'angle. Furthermore, both the object and background will be classified through calculating the two pixels—one from image and another from same location in estimated background image—vectors'angle. As for false margin and shadow, the paper suggests two effective methods of eliminating them respectively—varied square template and strip template. Experiments on given gait database demonstrate that binary object extracted by the new algorithm has fewer small holes and gaps and get good accuracy without shadow. Therefore, the segmentation effect using the algorithm is superior to that by background subtraction with morphology algorithm.(2) We put forward a type of method to descript 2-D gait contour: Tangent Vector Descriptor (TVD). Tangent vector is a effective way describing contour which is simple and intuitive. Firstly, gait contour be divided into equal intervals according the amount of contour points. Secondly for each interval we compute its midpoint's tangent vector by coordinate origin and interval endpoints as estimated value. Finally, the all midpoint's tangent vectors angle belonged to one contour make up a angle vector named Tangent Vector Descriptor (TVD). Experiments show that TVD could describe contour well.(3) We generate TVD class template from each key frame in very gait sequence. The Euclidean distance between each test sample's TVD and TVD class template is the similarity between the sample and the class.Nextly,a classical NN classifier be utilized to realize recognition. During previous steps high dimensionality feature vector be carried out eigenspace transformation based on the Principal Component Analysis(PCA).By fulfilling three aspects(gait images preprocessing, feature extraction and training recognition),this paper realizing a gait recognition system preliminary.