Search

This study proposes the design, implementation, and application of a new car-following data collecting method based on binocular stereo vision system. The car-following data consist of car-following distance, velocity, relative velocity and driving trajectory of a following vehicle. Specifically, triangulation principle is used to measure car-following distance. Robust locally weighted regression smoothing method is applied to automatically filter out the outliers caused by disturbance in real-world driving condition. Real-world experiments indicate that binocular cameras with 25mm focal length have advantages in traffic conditions with long car-following distance and high driving speed, and that binocular cameras with12mm focal length are suitable to measure on urban roads, where car-following distance is short and more driving conditions are needed to be observed. In 0-60m distance measuring range, it is verified that the mean absolute percentage error (MAPE) of binocular cameras with 25mm focal length is 3.2% and the MAPE of binocular cameras with 12mm focal length is 6.3%. The data collecting system helps to advance car following models due to its high accuracy and effective cost.

Bolt-loosening can cause poor running quality of trains, even resulting in terrible accidents. Currently, existing bolt-loosening detection methods for running trains need 3D data of the whole train body, extremely decreasing the efficiency of fault detection. In this paper, we propose a fast bolt-loosening detection method for the running train's key components based on binocular vision. Since a train generally consists of many cars with the same structure, the position distribution of train's key components is regular and periodic. First, we propose a novel method to detect key component regions including bolts, taking full advantage of this periodic distribution rule. Second, the sub-pixel edges of the bolt cap and mounting surface in the localized regions are extracted and segmented, respectively, combining with the convolutional neural network (CNN). Finally, based on stereo matching and the binocular vision model, the 3D data of these edges are obtained to calculate the distance between the bolt cap and mounting surface. By comparing the calculated distance with the reference value, we can judge whether bolt-loosening has occurred. The experimental results indicate that multi-bolt looseness can be calculated simultaneously. The measurement repeatability and precision are superior to 0.03 and 0.08 mm, respectively, and the relative error is less than 1.42%.

Localization of fruit and vegetable is of great significance to fruit and vegetable harvesting robots and even harvesting industries. However, uncontrollable factors, such as varying illumination, random occlusion, and various surface color and texture, constrain the localization of fruit and vegetable using the vision imaging technology under unconstructed environment. Our previous studies have developed various methods (illumination normalization, features-based classification, etc.) to localize a certain kind of fruit or vegetable using the binocular stereo vision. However, the localization of the multiple fruit and vegetable still faces challenges due to the uncontrollable factors. In order to address this issue, this study proposed an intelligent localization method of targets in fruit and vegetable images acquired by the two charge-coupled device (CCD) color cameras under unstructured environment. The method utilized the Faster region-based convolutional neural network (R-CNN) model to recognize the fruit and vegetable. Based on the recognition results, a window zooming method was proposed for the matching of the recognized target. Finally, the localization of the target was completed by calculating the three-dimensional coordinates of the matched target using the triangular measurement principle. The experimental results showed that the proposed method could be robust against the influences of varying illumination and occlusion, and the average accurate recognition rate was 96.33% under six different conditions. About 93.44% of 1036 pairs of tested targets from unoccluded and partially occluded conditions were successfully matched. Localization errors had no significant difference and they were less than 7.5 mm when the measuring distance was between 300 and 1600 mm under varying illumination and partially occluded conditions.