Train bogie part recognition with multi-object multi-template matching adaptive algorithm

Abstract Automation of train rolling stock monitoring system by recognizing the bogie parts is a process of identifying defects in the train undercarriage moving at >30 Kmph. Recognizing the parts of a moving train using computer vision models based on color and texture with deformable curve segmentation models is a challenging and computationally intensive. A multi object multi-template model is proposed to solve this problem in a computationally less intensive process. A multi object multi template library with 26 objects in 40 bogie frames was created based on statistical parameters of the bogie part in the center of the frame through template extraction model. Fifty templates were designed from 250 frames of the first bogie movement through the camera plane. Maximum normalized cross correlation coefficient calculated on each frame with a 26 – by – 40 template matrices identifies the bogie parts in the frame in a single computation. High speed recording of the train bogies at 240 fps establishes the datasets for experimentation having 2 trains with 20 coaches each capturing 15,000 frames per train. The correct recognition accuracy is 91% with a false recognition rate of 15%.