Multi-faced Object Recognition in an Image for Inventory Counting

Herein, an approach is presented to count objects in an image with different viewpoints. The multiple views of prototype have been used to get the viewpoint invariance in Speeded Up Robust Feature. For each view, presence of multiple instances in the scene image is investigated by matching the SURF features. If matches present, localization of instance is done in scene by generating a bounding box using homography. False matches are removed by computing correlation coefficient between transformed prototype and region of interest in scene. Different views of same prototype leads to multiple bounding boxes representing same object instance in scene. A supervised learning approach is used for classification of bounding boxes representing same instance. Hence, bounding boxes are grouped. Finally, a single bounding box which best describes an instance is chosen. The proposed algorithm is able to count objects for different viewpoints with better accuracy in multiple cases.