Research On Target Recognition Method Based Semi-supervised Learning

Research on object tracking and recognition is one of important research direction of computer vision field. It is an interdiscipline of computer science, image processing and electronic information science and has important application value on civilian and military fields. Target in the measuring system of shooting range is actually a common object. Effective recognition of a target is very significant for the performance testing and evaluation systems of the weapon equipments.A semi-supervised learning algorithm was designed and used in target recognizing and tracking in this paper. The algorithm is mainly composed of the two parts of tracking and detection. The main idea of the algorithm is taking the continuous target images acquired as a video sequence. The tracking and recognizing on the target are completed on the basis of taking the first frame information of video sequence as the template, The influence of the conditions, such as, camera shaking, illumination and scales changing on image etc. were effectively eliminated because of the detection is based on the semi-supervised learning.The basic principle on selecting feature points(corner points) was analyzed. Various features of optical flow methods as the tracking algorithms were studied and analyzed. Pyramid LK optical flow method was adopted in tracking object, which is effective in tracking large scale moving object.The application of classification and regression of random forests was discussed and studied. The decision trees which is usually used in random forests was not chosen in this paper, the random ferns was however used in this paper, which has the characteristics of simple structure, fast classification and less node threshold saving.The target tracking and recognition algorithm based semi-supervised was finally designed on the basis of the above analysis and discussion. The testing on human face, automobile and other video sequences were conducted. Satisfied results on the P(precision), R(recall) and F(comprehensive performance parameters) values were obtained. In target recognition, the overlap with target position and actual position is 90%, the recognition precision is100%. The algorithm can be used in various objects tracking and recognition.