The change of urban building is an important factor influencing urban development. It is particularly important for urban planners to efficiently and quickly understand building changes in the urban environment. However, most building monitoring operations still rely heavily on manual image recognition, which is not only time-consuming but also labor-intensive. Therefore, this study uses MS-FCN and U-Net deep learning models to assist in the detection of building change information in the Shezi Island area of Taipei City from true orthoimages in different periods. In the first stage of building recognition using MS-FCN deep learning model, the study added DSM (digital surface model) and DHM (digital height model) to explore the benefits of elevation on the model. The results of the building recognition stage show that adding elevation information from DSM and DHM can improve the model's building recognition ability compared to using only the aerial true orthoimages. The F1-scores achieved by adding DSM and DHM are 87.16% and 87.65%, respectively. In the building change detection stage, the U-Net deep learning model that was trained to resist registration errors and can achieve an F1-score of 71.63%. The results demonstrate the feasibility of using deep learning in combination high-resolution aerial true orthoimages and DHM to assist in building change detection operations.

In this research, we develop the stereo visual odometry based on the algorithm of the stepwise bundle adjustment, with mobile stereo camera system which is made by ourselves. Coplanar condition and Collinear condition, which are the concepts of the Photogrammetry, are utilized to eliminate the error matches after image matching. To improve the feature points, the concept of circular matching is added into the algorithm, which means keeping the feature points that is caught by all adjacent image pairs. The last step is using the stepwise bundle adjustment to solve the relative motion of adjacent image pairs, and we can rebuild the whole trajectory of the mobile stereo system with combining all the results. There are indoor scenario and outdoor scenario in the experiment of this research. Indoor scenario is the ground floor of department of Geomatics of NCKU; outdoor scenario is the square in front of museum of NCKU. The results show that the drift ratio of the two scenario are smaller than 1% and 1.6% respectively.

Autonomous vehicles have recently gained increasing popularity, and it is well-known that autonomous driving can reduce human error crashes and improve traffic safety. However, a few existing studies used two or more models to define the off-road environmental factors related to the crash severity level. Therefore, four research models (association rule, decision tree, random forest, and logistic regression) were used in this study for comparison. Our study data is California self-driving accident dataset from 2019 to 2021 (266 cases). The variables include self-driving car manufacturers, location, collision type, and severity. In addition, the number of various points of interest (POIS) near the crash location was also summarized based on the Open Street map. The results showed that factors such as if the crash involved an autonomous car from a tech-savvy company and minor vehicle damage, its severity level tends to be minor. Other related factors include shade in damaged area, movement preceding collision, and the density of POI (commercial, traffic).

The use of image processing for crack detection has become a common method in recent years, and drones are frequently employed today to assist with capturing images. By stitching these images together, a complete high-resolution image of cracks can be obtained. The stitching methods can be broadly categorized into traditional and learning-based approaches. Traditional methods rely heavily on selected feature points, and in scenes with fewer features or lower resolutions, stitching performance may degrade, potentially leading to stitching failures. In recent years, learning -based methods have gained popularity, harnessing the powerful feature extraction capabilities of Convolutional Neural Networks (CNNs). In this study, we adopted the multi-grid deep homography estimation network model proposed by Nie et al. (2021), which has depth-aware capabilities, for training specifically targeting cracks. Through experiments, we aim to identify the most suitable stitching method for continuous images of monotonous concrete backgrounds with cracks.