Auto-Calibration of Geolocation & Heading Direction: Segmentation based Matching between Network Camera & Satellite Images

Sammanfattning

Geospatial data is becoming increasingly important as artificial intelligence develops rapidly, creating new opportunities in the field of Geographic Information Systems. The thesis investigates the possibility of establishing a method to automatically calibrate a network camera’s geolocation and heading direction by using a camera view image and its corresponding satellite image. The aim is to locate the camera’s geolocation with one to two meters accuracy and its general heading direction. For the method, it is centered around four steps: Object-detection and segmentation, perspective transformation, cross-view image matching and the final calculations of the geolocation and heading direction. Three network cameras were used for the testing, all positioned in different locations and heading directions around an office building in Lund, Sweden. For the results, the method showed potential where the closest calculated geolocation was approximately 11 meters away from the actual position and around 30 degrees off for the heading direction. What was discovered during the thesis was that e.g. none of the cross-view matching models used were ideal for the thesis purposes. In conclusion, the method has four distinct steps which makes it hard to identify potential local error sources, however, when one is located the following steps improve accordingly. Cross-view matching was deemed the step which had the most influence over the final results and the one needed to improve the most to gain a more accurate geolocation and heading direction of a network camera.