Native Vehicles Classification on Bangladeshi Roads Using CNN with Transfer Learning
Shaira Tabassum, Md. Sabbir Ullah, Nakib Hossain Al-nur, Swakkhar Shatabda · 2020 IEEE Region 10 Symposium (TENSYMP) · 2020
To bring an intelligent transportation system in a developing country like Bangladesh, vehicle detection or identification is a major concern to analyze the unique behaviors of the native vehicles. Coming with the tremendous progress in object detection with deep convolutional neural network (CNN), we have addressed the recognition of native vehicles on Bangladeshi Roads. In this paper, a transfer learning approach based on CNN has been applied to the popular You Only Look Once (YOLO) framework for vehicle classification. The goal is to reuse the pre-trained convolutional layers of the YOLO framework and transfer knowledge to detect 15 novel vehicles of Bangladesh. To validate our proposed method, we have introduced Bangladeshi Vehicles Dataset containing 9000 annotated images of 15 vehicles. The images are extracted from Bangladeshi traffic scenes. Experimental results demonstrate that the proposed method has achieved an admissible outcome with 73 % IoU to properly detect native vehicles. Thus, the system can progress the traffic management system of developing countries.