Understanding buildings in agricultural area is important because the arable land in Taiwan is limited. One of the practical ways is manual digitization from high resolution satellite imagery, which can acquire satisfying results without field investigation. However, such practice is tedious and labour intensive. Given these reasons, past research devoted to deep learning approaches have shown that convolutional neural networks are useful for building segmentation using satellite imagery. In this study, ENVINet5 model was trained and utilized from high resolution Pleiades pansharpened imagery. The training images (with the size of 2500 pixels × 2500 pixels) were randomly selected from 9 counties/cities to increase diversity because each county/city has different building patterns. The performance of ENVINet5 model was evaluated based on pixels and polygons, respectively. The pixel-based evaluation showed that the trained model can find 84% of building pixels. The polygon-based evaluation was carried out through calculating the number of building segments and comparing them with the reference data using IoU (Intersection of Union). The results showed that 92% of building segments were found, and the IoU of most building segments range between 0.6 and 0.9. The trained model was validated on the testing images for the transferability test. Moreover, an image tiling and stitching technique was proposed to deal with large satellite imagery. Finally, we compared the time costs of labelling with and without the aid of deep learning approach. The results showed that the time costs decreased by 7.3% with the help of deep learning approach.

The ongoing race toward an autonomous era results in the development of High Definition (HD) Maps. To help extend the vision of self-driving vehicles and guarantee safety, HD maps provide detailed information about on-road environments with precise location and semantic meaning. However, one main challenge when making such a map is that it requires a massive amount of manual annotation, which is time-consuming and laborious. As such, to fulfill automation in extracting information from the sheer amount of data collected by mobile LiDAR scanners and cameras is at most concern. In this study, a workflow for automatically building traffic sign HD maps is proposed. First, traffic islands, traffic signs, signals, and poles are extracted from LiDAR point clouds using PointNet. Then, point clouds of traffic signs are clustered by the DBSCAN algorithm so that the geometric information can be obtained. An evaluation is performed to assess the accuracy of geolocation in the final stage. Next, point clouds in each traffic sign cluster are projected onto corresponding MMS images for classification purposes. The semantic attribute is obtained based on the GoogLeNet classifier and determined by a proposed mechanism, i.e. modified SNR ratio, which ensures the class with the most classified images is significant enough for that cluster to be considered as that specific type. An output text file including precise coordinates of traffic sign center, bottom-left, and top-right of the traffic sign bounding box also their type is generated for further use in HD maps.

The feature-based Visual Simultaneous Localization and Mapping (V-SLAM) systems localize cameras by repeatly establishing the 3D feature maps, matching the image features to the 3D feature maps and conducting space resection. By reusing the 3D feature maps established during SLAM processing, it can be considered as a control field to solve the GNSS-denied localization problem. However, the different lighting condition in the outdoor environments can possibly affect the results of image feature extraction and matching, which will lead to different localization results. This study applies ORB-SLAM to establish the 3D feature maps, and the map-aided camera localization is conducted under ORB-SLAM Localization Mode. The experimental results demonstrate the camera localization aided by a 3D feature map can reach close camera localization accuracy as the accuracy of its’ reference 3D feature maps. As a result, the camera localization aided by a 3D feature map can become a potential solution of GNSS-denied localization problem.

3D building model is one of the important elements in digital city analysis. It can be widely used in many geographic activities, such as smart city, urban planning, disaster management. The quality of 3D building models is related to their structure and geometry. CityGML, an international 3D city modeling standard, formulated the scale to express the detail of 3D building models in different Level-of-Details (LODs). From the roughest to the most detailed one is named LOD-0 to LOD-4. The main goal of our study is to reconstruct the LOD-2 3D building models which are configured by detailed roof structure with vertical facades. Since the 3D roof structure is the most important part of the LOD-2 model, we will focus on how to reconstruct the complete high-accuracy and topological error-free 3D roof models. We create point clouds and Object Height Model (OHM) from Unmanned Aerial Vehicle (UAV) images and Airborne Laser Scanning (ALS), extract 2D polygons of the roof structure by manual digitization, then perform least-squares adjustment to fit the 3D roof planes. Considering the geometry and relationship between adjacent polygons, we set some constraint conditions and conduct roof plane correction to avoid topological errors. Eventually, we reconstruct some typical roof models in Taiwan, and conduct accuracy analysis by manually measuring the 3D coordinates of roof corners. It has proved that the results can conform to the LOD-2 standard of CityGML, confirming that the proposed method of 3D roof models reconstruction is feasible to varied types of roofs with high accuracy.