Recently, the railway transportation is experiencing high growth rates all over the world, especially, in China. However, the high-speed railway faces more serious security problems than traditional railway, and railway maintenance has become a more and more important issue in railway development. Missing bolts detection is one of several key issues on railway maintenance. Bolts are usually used to fix the rails to the sleepers. Thus, it may lead to the serious accidents due to the absence of blots. In order to tackle this problem, this paper presents a bolts detection algorithm to improve the performance and efficiency based on our previously established system.Our previously established system has achieved the basic functions. However, it still has some shortcomings and should be improved in both performance and efficiency. The main challenges are as follows:(1) First of all, the inaccurate position of track may cause inaccurate search range of bolts. It may lead to inaccurate bolt detection results. Thus, improving the performance of the track position becomes a key issue.(2) The detection speed of our previous system is somehow slow. Therefore, we will focus on improving the efficiency.The main content of this paper are as follows:(1) Firstly, we need locate the bolts correctly. Fortunately, the bolts are always close to the rail and the distance between rail and bolts is constant. Therefore, if we can predict the location of the rail, the search region that may contain the possible bolts is extremely reduced. In our previous system, we used an unsupervised algorithm to detect the rail track. However, there are some constrains might be more helpful, such as the temporal information between the consequent images. Therefore, in this thesis, we propose a novel supervised algorithm to verify the detecting results. Firstly, we obtain the standard deviation range of the track position, which can be used to verify the detecting result of the previous unsupervised algorithm. If a previous result is not within the standard deviation range, we will correct the result.(2) We apply the HOG characteristics and the nearest neighbor algorithm based on Hamming distance to bolt recognition in the track images. HOG descriptor is used to describe the gradient and edge characteristic of the objects, it is a good descriptor of the shape characteristics of a bolt. The nearest neighbor algorithm based on Chi-square distance is applied to the previous system. However, bolt detection based on HOG needs exhaustive search which leads to the low efficiency. Thus we propose the nearest neighbor algorithm based on Hamming distance. Hamming distance is a metric function which compares the corresponding bits in two feature histograms. Its calculating speed is pretty high due to the bit operation. The experimental results demonstrate that the nearest neighbor algorithm based on Hamming distance can improve the efficiency of bolt detection in the track images. |