Vehicle Detection from Ultra-high Resolution Aerial Image, Three Line Scanner

Year by year, traffic congestion is severe on street in big cities such as Tokyo because vehicles increase rapidly compared to constant street surface. Furthermore, this problem causes economic and environmental waste. Therefore the attempt to monitor vehicles and their behaviors becomes challenging demanding unavoidably on vehicle statistical analysis for efficient and effective traffic planning and management. One potential method of several vehicle-monitoring ways is on aerial images with digital photogrammetry algorithms due to observation of wide-area coverage at the time. Recently, new aerial sensor, Three Line Scanner (TLS), has been realized with a capability of centimeter-ordering resolution and seamless imaging digitally as well as linear object trackability. TLS images are well fitted to fulfill Vehicle Monitoring on streets in Mega City. Therefore, unique and original algorithm of Vehicle Detection from TLS image has been researched and realized. The framework of this study is divided into three steps: Preparation Step, the Automatic Vehicle Detection and the Semi-Automatic Vehicle Detection in difficult area. In preparation step, at first stage, the raw TLS image is segmented by region growing At later, noise removal and over-segmented region correction are by morphological techniques and geometric constrains as well as nearest interpolation. Also, a nearest-region network is generated by dilation techniques. Building Shadow, detection occlusion, is detected and corrected by sunny-shadow object relation and overlay techniques. Moreover, the road, regions of interest for vehicle detect, is extracted as linear-feature objects by radon transformation and buffering methods. At a final proceed of preparation step, regions on street is detected. Non-vehicle regions as road surface, lane line, road mask are filtered and only rest of vehicle region candidates are still existed. In automatic vehicle detection stage, the vehicle model hypothesis has been crated from generic vehicle component and vehicle size catalog. At time, Vehicle Shadow/ Dark Vehicle is detected and applied as vehicle cue. At a cue position, “cluster” neighbored regions along sun side is detected by frame detector. “Cluster” is merged as a region by expansion process. Finally a “cluster” as vehicle, satisfied with a vehicle model hypothesis, is detected. Some vehicles especially under occlusion such as a building shadow, however, are not detected. Therefore, at the semi-automatic vehicle detection, some region is given as the starting point to generate “cluster” by the expansion process. A “cluster” as vehicle is validated and matched with the vehicle model as same processing in a automatic vehicle detection. Finally, the robust method has been developed and promising results are performed in this paper.