| dc.description.abstract | This thesis proposes a drone based automatic inspection system for illegal parking. Using this system, traffic policeman sets the route to be patrolled, and then the drone can fly autonomously following the set patrol route and automatically detect illegally parked vehicles on the roadside. When an illegally parked vehicle is found, the drone will lower the altitude to take its license plate and record the information of the violation, then the photos and information will be sent back to the policeman traffic control center. The entire process is automatically performed without any personnel operation.
The system’s function consists of three parts. The first part is to obtain Global Positioning System (GPS) coordinates for the set patrol route and to control the drone flying along the route for detecting illegally parked vehicles. The second part is to identify the vehicle with parking violation. Four common parking violations in Taiwan, side-by-side parking, red-line parking, reverse parking, and grid-line parking can be automatically identified whether the illegally parked vehicle is a car or a motorcycle. The third part is to capture the license plate of the illegally parked vehicle and to recognize the license plate number and other illegal information records. When an illegally parked vehicle is found, the drone will fly lower to the back of the vehicle and use the object detection network to detect the license plate and recognize characters on the license plate.
This thesis uses YOLOv5 to detect and recognize ehicles, grid-lines, license plates and license plate numbers; applies Hough line transformation to obtain the position of the road board in the image; adopts EfficientNet classification network to identify the direction of the vehicle; proposes a loss function that combines classification and regression errors to improve the accuracy of the training of EfficientNet; and utilizes fuzzy control technique to control the flight attitude of the drone for correctly flying and photo taking. Compared with traditional manpower patrols, this system indeed greatly reduces the burden of the traffic policeman on duty. In the real road experiments, the system has an identification accuracy of 92.13% in four violation situations. Moreover, the flight control of the drone can overcome the influence of winds below grade 5. The overall experiment was very successful.
Keywords: Drones, fuzzy control, parking violation detection, deep learning | en_US |