| Abstract [eng] |
In this work is tested and compared four different object detection methods. Objects selected for testing has different shape, color, size. At the beginning training images is collected. To increase size of training data, augmentation is used. One hundred images samples are collected for each test object. Sample images can be split in two categories one with natural light and no additional objects. Second with artificial light and a few additional objects in testing background. Comparison criteria to evaluate methods are average precision, mean square error, average detection time, percent of successfully recognized objects. Criteria values calculated from the object detection methods results in pixel coordinates. All these coordinates are compared with ground truth object positions in testing images. Two different methods used to convert pixels coordinates to object positions in real world measurements. One method is based on depth camera data, another on RGB image, markers, and proportions. Both methods have similar error values. Calculation based on depth camera is simpler to use and gives some additional info about object. All results from object detection methods are summarized and compared at the last chapter. |
