Improving the performance of a SVM+HOG classifier for detection and tracking of wagon components by using geometric constraints

Abstract

The inspection of train and railway components that can cause derailment plays a key role in rail maintenance. To improve productivity and safety, service providers look for automatic and reliable inspection solutions. Although automatic inspection based on computer vision is a standard concept, such an application challenges development community due to the environmental and logistic factors to be considered. Previous publications presented automatic classifiers to evaluate integrity and placement of wagon components. Although the high classification accuracy reported, ineffective object detection affected the general performance. Our object detector/tracker consists of a descriptor based on the histogram of oriented gradients, a support vector machine classifier, and a set of geometric constraints, which takes in account the ideal trajectory path of the wagon’s components of interest and the distances between them. We detail training and validation procedures, together with the metrics used to assess the performance of the system. Presented results compare two other techniques with our approach, which exhibits a fair trade-off between reliability and computational complexity for the application of wagon component detection.