Performance of Traffic Sign Detection & Warning System
Traffic signs are vital in ensuring the proper running of traffic on our roads. However, there are still a great deal of accidents occurring due to ignorance or failure to pay attention to the signs. A traffic sign detection and recognition system is therefore developed as an initiative to diminish the rate of accidents. The proposed system is customised for the people of Mauritius and it allows a language choice between French and English. Moreover, the application also supports learning of traffic signs since it is embedded with a quiz. Several methods were consulted through literature review to find the best one that will suit our system. After analysing each one of them, the method chosen was colour and shape detection to localise traffic signs from pictures and video frames and Convolutional Neural Network to classify the detected signs accordingly. The dataset used for training was constructed by merging the German dataset with local traffic sign classes. Multiple occluded traffic sign images were included in the dataset since there are many partially obstructed traffic signs on our roads. An accuracy of 97% was obtained after training the classifier with the custom dataset. The system allows the user to classify an uploaded image or to classify signs in real time. It outputs the name of the traffic sign together with the classification probability. The application is integrated with a user-friendly interface. A vocal notification feature was also added to make it more practical. The system is able to classify signs in different lighting conditions by using a brightness adjuster, it can classify multiple signs in an image, with varying distances and inclination and lastly the system can also classify occluded signs correctly.
Completed by:
Mandary Ganish
BSc (Hons) Computer Science 2022
