การจำแนกประเภทวัตถุโดยการใช้ความสัมพันธ์เชิงพื้นที่และการปรากฏร่วมกันกับปฏิสัมพันธ์ของคน

• Main Author: ประภัสสร วิสุทธิรังษีอุไร
• Other Authors: กานต์ ปทานุคม
• Format: Theses and Dissertations
• Published: เชียงใหม่ : บัณฑิตวิทยาลัย มหาวิทยาลัยเชียงใหม่ 2020
• Subjects: ความสัมพันธ์เชิงพื้นที่; ปฏิสัมพันธ์ของคน
• Online Access: http://cmuir.cmu.ac.th/jspui/handle/6653943832/69414
• Institution: Chiang Mai University

The object categorization or recognition is common problem in computer vision that can be applied in many applications such as in surveillance cameras or intelligent robot system. There are many researches on the efficient appearance features for categorization systems, however, in some practical situations, the efficiency of the appearance based features may be decreased if there are high intra-class variation in the object categories or the deformable objects or occluded objects. In this research, we propose a human interaction based framework for object categorization. In the proposed framework, we improve the object categorization system by applying co-occurrence and spatial relationship based features. We focuses on an interaction between human and object in the indoor scene. The videos that human are performing the action of using the object are used as inputs and training samples of the proposed system. Co-occurrence of objects and hand postures, a relative position between objects and face, an object motion, and an object appearance are concatenated and used as the features. The co-occurrence based features are defined as co-occurrence between the hand posture prototypes and objects. The hand posture prototypes can be obtained by -means clustering. The histogram of relative positions between object and face and histogram of object motion vectors are applied as two features of spatial relationship based features. In the experiment, we evaluate by using 108 videos of six classes of objects. The results show that the proposed framework can provides the accuracy of 89.63 % which improve by 27.59% in comparison with the appearance based feature.