An Adaptive GPGPU-based Bolt Detection System

Railway bolts detection is an important task of railway maintainance. In the historical background of rapid expansion of high speed train, railway maintainance becomes more and more important. As a subtask of railway inspection, automatical bolts detection becomes an increasingly important research subject. This thesis provides some improvements on performance and effieciency.The main contributions of this thesis are as follows:1. Since there’s no bench mark for bolts detection, we build a semi-automatic manual bolts calibration tool to create it. The bench mark contains5railway files, including13,416images,80,026bolts, which includes79,043normal bolts and983missing or contaminated bolts. We also provide an automatical experiment evaluation system which can evaluate the scores automatically according to the experiment result and bench mark. So it can significantly shorten the experiment time.2. We propose an online learning based self-adaptive bolt detection algorithm. This method prevents the threshold in the nearest neighbor classifier that used in the previous system and can automatically adapt to a new railway line or a new recorded video. It achieves the accuracy of91%～99%for different railway lines in the experiments of this paper.3. In order to meet the requirement of realtime in bolt detection, we adopt the CPU-GPGPU (General-purpose computing on graphics processing units) heterogenes method to parallelize several most time consuming sub algorithms, and collaborate CPU and GPU to achieve higher computing effieciency. The speedup of HoG descriptor and distance calculations in the experiment reaches17.73and11.22respectively, and the whole system get the speedup of6.11.