A Transfer Learning Approach for Identification of Two-Wheeler Brand and Model

Abstract

Motorcycles were originally designed to provide safer, more efficient, and more comfortable rides, but they are now also used for criminal purposes. Due to overspeeding, motorcycles often get involved in accidents, and it is challenging for law enforcement officials to identify the culprits from CCTV footage or spectator accounts. This paper presents a solution to this challenge by using a pre-trained deep learning model to detect, classify, and identify motorcycle models. To overcome the limited availability of annotated bike databases, this proposed work created a new bike dataset that includes 5000 annotated images sourced from major search engines, CCTV footage, and manual captures from bike showrooms. Then the bike brand was identified in 27 classes by the Faster RCNN pre-trained model and achieved an accuracy of 94.35%. The proposed model was compared with the other pre-trained models such as YOLO V5 and MobileDET, among these, the Faster RCNN provided better identification accuracy.