The Study Of Structured Light Based Railway Turnouts Recognition

Turnout is not only an important part of railway track, but also the very part with more wear and more impact. Therefore, regular detection of the parameters and evaluation for the working state of turnouts are needed. As is known, the evaluation standard for turnouts is different from that of ordinary tracks. However,present track inspection car can't recognize turnouts from detected objects. Therefore, in the process of evaluating track state by human,we need to distinguish turnout's data from ordinary track's data. On the other hand, without fully taking advantage of the characteristics of tracks, responder-based train positioning method is facing the contradiction between capital and density limit of responders. As an inherent part of railway track, turnouts can be used as an assistant method of train positioning, leading to less dependency on responders. Moreover, it is a necessary step to distinguish turnout's curves before calculating turnouts wear. The three situations mentioned above arouse an important question. That is, how to automatically classify turnouts' figures and normal railway track figures.Researches on turnout recognition have being rare .This thesis explored the methods for automatic turnouts recognition.After analyzing the features of turnout shape curves, we give three solutions. In solution one, we use an vector to represent a curve, and then calculate the distance between two vectors. And last use a comprehensive procedure to classify curves. But it failed. In solution two, a Relative Height Based Curve Description algorithm is used to code curves into a vector, which is later imported into a decision tree model called CART. In solution three, we created a simple Convolutional Neural Network and trained it with track curve data as train patterns. As was indicated in the results of solution two, a combination of Relative Height Based Curve Description algorithm and CART decision tree model worked well with 87% of turnout figures correctly classified. However, solution three failed with no turnouts figures correctly classified because the turnouts data quantity is so limited which causes an unbalance between train data and test data.