Abstract:
Detection of vehicles and its license number plates is an important part in real-time applications. Many of the approaches initially detect a vehicle and number plate, after that recognize the characters by using only an image of the specific vehicle. This approach has low recognition rate because of the noise present in that specific frame. However, in our proposed system a different real-time approach is used to detect the vehicle and localize its license number plate instead of selecting a single frame to perform the recognition. In live video stream vehicle was detected, then by applying Laplacian filter having a certain threshold value, clear frames were obtained for further processing. Many of the existing solutions are not fast for real-world situations because of several constraints. In this thesis we present a robust and efficient vehicle detection and license number plate recognition system based on the state-of-the-art SSD object detection. By using transfer learning a new model was trained based on the collected dataset using SSD (coco dataset) pretrained model. This dataset contains 1693 images having 800x600 resolution also vehicles are of different types (cars, motorcycles, buses, and trucks). The SSD and KNN models are trained and finetuned for each stage of vehicle detection, license plate localization, character recognition and Color detection, so the system is robust under different conditions i.e. due to camera, lighting or background. The extracted information of vehicle number plate, type, color, time, speed, location, cropped images of vehicle and number plate will be stored in the database created using MongoDB. Also, specific data can be accessed from database using search option on frontend GUI application. The success of
our system greatly depends on the quality of acquired video. Our real-time system efficiently and successfully localizes license number plate under different environmental conditions i.e. indoor, outdoor, or daytime. It has ability to detect license number plates of different provinces. These detected number plates can be of different colors, having different fonts, number plates may be having solid color background or an image as background. Depending upon all the mentioned factors accuracy of vehicle detection is 97.23 % and number plate localization is 89.84%. The live video stream is having 30 fps for which it takes 15ms to detect a vehicle.
