In recent years, the shortage of truck drivers in domestic logistics has become apparent. One solution to this problem is the use of articulated vehicles. An articulated vehicle has a large capacity trailer, which can increase the amount of transportation per driver and help meet the transportation demands of domestic logistics. However, it requires a high level skill to drive an articulated vehicle. In particular, reverse parking of articulated vehicles requires different steering wheel maneuvers than parking of normal cars. Automation is expected to assist inexperienced drivers to operate reverse parking. In this study, we verified whether it is possible to acquire autonomous reverse parking capability using deep learning. We applied two types of deep learning methods, i.e. with and without human driving data. We used Deep Deterministic Policy Gradient (DDPG) for the method without data, and Generative Adversarial Network (GAN) for the one with data. We verified that it is possible to acquire a certain level of autonomous reverse parking capability through learning by DDPG, although it is not as good as learning using human driving data. We also verified that it is possible to acquire autonomous reverse parking capability similar to human operation through learning by GAN.

The purpose of this study is to develop an in-line inspection technique for detecting fine defects on the surface of metal cylinders. The magic-mirror method, which involves a wide field of view and exhibits high sensitivity, was employed, and a prototype inspection system was developed. This system operates as follows: a sheet beam with a width of 350 mm is irradiated onto the surface of the sample, and the reflected light illuminated onto a cylindrical screen is captured through a line sensor camera; the screen can be rotated to reduce speckle noise. When the cylindrical screen was rotated with a peripheral velocity of 883 mm/s, the speckle contrast decreased by 36% compared to the case when the screen was non-rotating. To evaluate the detection sensitivity, the surface of a metal rod with a diameter of 22 mm and length of 340 mm was inspected. Experimental results show that the system is able to detect concave defects with a diameter of 35 µm and depth of 3.9 µm as well as hairline defects with a width of 75 µm and depth of 1 µm as light and dark patterns on the inspection image.

Digital image has become an important data representation in precision engineering due to recent advances in image acquisition and machine learning technologies. The contents of a digital image consists of various scales, and therefore scale-aware image filter, which removes small structures while preserving salient features, is useful in many applications, but is computationally expensive in practice. This paper proposes a simple and fast computational method for scale-aware image filtering based on half-box regions. The proposed method consists of a recursive process of joint bilateral color averaging approximated by assuming constant bilateral weights within each half-box region adjacent to a given pixel. These bilateral weights are designed to also remain constant during the recursion process. Therefore, they only need to be calculated once, which reduces the computational cost dramatically. The color sums within the half-box regions are efficiently calculated by applying only one fast box filter using the relative coordinate relationship between the pixel and each half-box center. The quality, speed, and convergence rate of the proposed method were also examined by numerically comparing it with conventional methods, and it was fond that the proposed method achieved high performance.