Developing an automated tree localization system through street-level image analysis

• Main Author: Pratama, Leonardo Irvin
• Other Authors: Lee Bu Sung, Francis
• Format: Final Year Project
• Language: English
• Published: Nanyang Technological University 2023
• Subjects: Engineering::Computer science and engineering
• Online Access: https://hdl.handle.net/10356/165928
• Institution: Nanyang Technological University

In this project, I aimed to develop a system for localizing trees in images collected from Mapillary. The ability to accurately identify and locate trees is important in a variety of fields, such as ecology and forestry, where knowledge of the distribution and species of trees can provide valuable information. To achieve this goal, a combination of object detection algorithms and mathematical concepts was employed. The first step in the approach was to collect and label a dataset of tree images from sources like Mapillary, Singapore Land Authority, and self-taken images. This dataset was used to train the object detection algorithms, namely YOLOv7, which were designed to learn features such as bark texture, trunks, and overall tree structure. This process involved several rounds of augmentation, training, and testing to optimize the performance of the models, whose mAP@0.5 went as accurate as 0.8578. Once the models were trained, they were used to detect bounding boxes of trees from the images in the dataset. These boxes and their image metadata were used as inputs to the localization algorithm. The object detection algorithm, which is a product of YOLOv7 framework, was used to identify and bound the trees in the images. The algorithm processes the images in multiple scales and detects the trees by recognizing the features learned by the convolutional layers. This stage is subsequently completed with the introduced localization system, which utilizes triangulation to extract metadata from these images and locate objects within the detected objects. The system also provides the coordinates of the bounding boxes around the trees, which allows for geo-localization of the trees within the image. To evaluate the performance of the system, the system was tested on a separate test set of images. The results showed that the system achieved high accuracy in both tree localization and recognition. The model achieved a precision of 0.9483. On the other hand, the algorithm was able to accurately bound the trees in the images with a high degree of precision. Subsequently, the tree distribution was able to be plotted within the specified area with a distance mean square error of 6.33 metres. In conclusion, the system developed in this project proposes an algorithm to determine the tree distribution within an area using publicly available dataset provided by Mapillary. The high accuracy of the system in both localization and recognition suggests that it can be applied in a variety of fields, such as ecology and forestry, where accurate identification and location of trees are important. Further research could be conducted to improve the performance of the system and to explore other potential applications.