A heuristic approach to screw pose detection

Sammanfattning: Object detection using vision systems could generally be accomplished using machine learning or conventional computer vision techniques. This thesis implements and compares three methods to locate screws on a flat plane using conventional computer vision on regular non stereo images without any depth or range information. A framework based on OpenCV and C++ were developed to aid testing of common image processing techniques to find suitable processing pipelines for screw detection. Two edge-based techniques in which screws were modelled as parallel lines of some length or as bounded regions with some area and one naive template match based technique were compared. Edges were found using the Canny edge detection algorithm, lines were extracted using the Hough transform and contours were extracted with a border following algorithm. The template-based method brute-force tested rotated templates in parallel. Machine learning would generally require large datasets to be able to detect screws. The conventional methods presented in this thesis instead requires some manual parameter adjustments depending on the screws type and the surrounding environment. These parameters could be adjusted once for a certain combination of screws and environment and be left untouched until some change is introduced in the system. The algorithms were tested on a set of images from 6 different combinations of screws and backgrounds. Every combination was tested 3 times and the screw locations were randomized between every test. The line-based model successfully located on average 65% of all visible screws in the test images, while the contour and template matching models were able to detect 53% and 61% respectively. The two edge-based methods had a higher error rate, but 30 to 60 times faster processing time than the naive template matching based method. No information about the screws orientation (z-axis rotation) or distance from camera (z-axis position) could reliably be extracted. Future studies could look into the possibilities in how these three different methods could be combined within themselves or with other techniques to extract orientation and depth data.