| Abstract [eng] |
Image recognition and processing technology development takes a relatively large part of all new technology development processes. About 90% of all information received from surrounding environment a person receives as a visual information, so it is obvious that the integration of video recognition technology will be widely developed. The range of use is quite large, ranging from security equipment to the autonomous vehicles, which are becoming a huge interest of the car manufacturing companies. My aim was to write a program for image recognition using Raspberry Pi 2 mini-computer, which can recognize objects, detect an obstacle, to go to the desired position and to control mechanical devices according to visual recognition results. Artificial neural networks were used to increase the versatility of the program in different situations. Different image recognition models were tested: like moving objects detecting algorithm which uses optic flow principle, obstacle and road surface detection program, facial recognition software, SURF visual recognition program and image recognition programs with Hough transformation algorithms. During the time of video recognition experiments it was found, that the response time of video recognition program with Hough transformation was in average 0,6 of a second, which means that fast moving objects would not be detected correctly and image processing will falter and will be unstable. Neural network adaptability and the ability to learn is very useful in a visual object recognition. For example, when you need to look for similar objects - such as faces. Night vision system eliminates colour deviations due to elimination of shadows and uneven lighting. Neural network image recognition accuracy depends on the amount of training data and the number of training cycles – it means it can run fast, but not as precise, or more slowly but more accurately. SURF algorithm is suited for complex object recognition, it is resistant to uneven lighting and can recognise even deformed and rotated objects, but the Raspberry Pi minicomputer is too slow to use it for a real-time video recognition. |
