Current state-of-the-art intensity and depth modal-based railway surface defect inspection system faces the dilemma between false alarming and miss detection. To overcome this challenge, a bimodal detection scheme using both feature-level fusion and (evidence theory-based) decision-level fusion is designed. Moreover, an improved evidential fusion algorithm is proposed, which adopts three-branched evidential weight structure and introduces the Transferable Belief Model to the improved decision functions, achieving outstanding performance on effectiveness and robustness analysis. Both theoretical justifications and experimental results validate the efficiency of the hybrid fusion-based detection scheme in solving rail surface defect problems.

We propose a monocular visual odometry method for railway localization, FRVO-Mono, which exploits different features in railway environments, such as vanishing points, track lines, and pole objects to provide multi-dimensional geometric constraints for train location estimation. A multi-view measurement function between the camera, points, lines, and objects is constructed to correct rotation and scale drift. A 3-step object association strategy is also suggested to improve objects matching accuracy, and the prior information of the camera’s height is used to help restore the absolute scale. Data from four real railway scenarios with different scales and speeds are collected in collaboration with China Academy of Railway Sciences (CARS). The experimental results demonstrate the great accuracy and robustness of our approach, as well as its ability to effectively rectify the scale drift to meter level within a given distance, which makes it possible to generate more accurate positioning when combined with some other semantic information.

In current smart classroom research, numerous studies focus on recognizing hand-raising, but few analyze the movements to interpret students’ intentions. This limitation hinders teachers from utilizing this information to enhance the effectiveness of smart classroom teaching. Assistive teaching methods, including robotic and artificial intelligence teaching, require smart classroom systems to both recognize and thoroughly analyze hand-raising movements. This detailed analysis enables systems to provide targeted guidance based on students’ hand-raising behavior. This study proposes a morphology-based analysis method to innovatively convert students’ skeleton key point data into several one-dimensional time series. By analyzing these time series, this method offers a more detailed analysis of student hand-raising behavior, addressing the limitations of deep learning methods that cannot compare classroom hand-raising enthusiasm or establish a detailed database of such behavior. This method primarily utilizes a neural network to obtain students’ skeleton estimation results, which are then converted into time series of several variables using the morphology-based analysis method. The YOLOX and HrNet models were employed to obtain the skeleton estimation results; YOLOX is an object detection model, while HrNet is a skeleton estimation model. This method successfully recognizes hand-raising actions and provides a detailed analysis of their speed and amplitude, effectively supplementing the coarse recognition capabilities of neural networks. The effectiveness of this method has been validated through experiments.