Study On The Recognition Algorithm Of Lunar Hole Based On Machine Vision

The moon is the only secondary planet of the earth, and is the nearest celestial body of the earth. So the moon exploration has great significance in the scientific, economic, military and political fields. China has carried out its first plan of lunar exploration"Chang E"Project on March 1st, 2003. This project is carried out through three steps; they are circling, landing and returning. The Intelligent Vehicle Group of Jilin University have been making great efforts to Lunar Rover design and lunar environment perception, now have make great performances. The positive and negative parallelogram Lunar Rover which has independent intellectual property, synthesizes image processing, environment perception, multi-sensor information fusion, intelligent automatic navigation, path plan, obstacle avoidance and such technologies.The research of Lunar Rover is in the need of the development of lunar exploration. As the tool to explore the moon, Lunar Rover is an important component of the project. Lunar Rover is a robot combined with remote control and autonomous navigation. As a pioneer landing on the moon, it can realize long-term practical exploration in stead of the astronaut and can bring back experimental samples that can be analyzed by researchers. The Lunar Rover is a robot that combines many techniques such as mechanical design, computer science, electronics, communication technology, materials science, autonomous control and so on.There are a lot of complicated holes of different sizes and different ages in the moon. This brings great difficulties for the navigation of the Lunar Rover. So it is the key technology to detect the location and size of the holes for the Lunar Rover's autonomous navigation and smooth steering. A period of time research shows that it is feasible to detect whether there are holes in front of the Lunar Rover by using machine vision.In order to make some technology support and project authentication for the navigation method of future Chinese lunar rover, cooperated with China Academy of Space Technology, the feasibility of the perception of lunar environments using machine vision is discussed.Based on the above reasons, the paper mainly consists of four parts:1. Pre-processing algorithms of lunar hole images are studied. Median filtering method, average filtering method, histogram equalization and edge enhancement method are introduced. Location of lunar hole is obtained by using horizontal and vertical projection method. The segmentation of lunar hole in faintish illumination is realized by using edge intensity method and Fisher Criterion method. After comparison, the Fisher Criterion method is selected. Experimental results show that the algorithm is effective and reliable.2. Algorithm of segmenting the lunar hole in normal illumination is studied by using region growing method and two-dimension maximum entropy method. Researching on selecting the seeds, establishing the growing rules and confirming the time when growing is over is carried out according to the principle of the region growing algorithm. Because one-dimension gray-level histogram can only reflect the gray distribution of the image, the correlative information of pixels is not described. When the image is complicated or the illumination is changed, the segmentation result is not satisfactory. Because the relativity of pixel and its neighbors is very big, the distribution of the object and the background is much easier to be distinguished in the two-dimension histogram than in the one-dimension histogram. This paper found a two-dimension histogram using the gray level and its neighbor pixels average gray level, thus the accuracy and the immunity of the segmentation can be improved obviously. After comparison of experimental results, the two-dimension maximum entropy method is selected to segment the lunar hole in normal illumination.3. Method of segmenting the lunar hole in strong illumination is studied by using moment invariants algorithm and OTSU method. First introduces the principle of the moment invariants algorithm: an image which has only the object and the background will become a binary image after processing. The original image and the binary image both have its moment respectively. The moments will be adjacent or equal if the threshold is good. The moment invariants algorithm is based on the invariability of the first three moments of the two images. The OTSU method can divide an image into two parts based on the gray level. The bigger the variance between the clusters is; the bigger the difference between the object and background is. If there are some mistakes, the variance will become smaller. So the maximum variance means the minimum error. After comparison, the moment invariants algorithm is selected.4. The method of using neural network to classify three kinds of lunar hole images in different illumination is studied. The neural network classifier of lunar hole image is brought forward, and structure of the neural network is designed. The typical data of different lunar hole images is used to train the classification network. The experimental results show that the neural network classifier can efficiently distinguish different lunar holes images.5. All the algorithms are integrated into a system-software on Visual C++ platform. The related experiments are carried out. Experimental results show that the algorithms have good performance and good robustness.