Improved Vehicle Logo Detection and Recognition for Complex Traffic Environments Using Deep Learning Based Unwarping of Extracted Logo Regions in Varying Angles

Abstract

AbstractIntelligent Traffic Monitoring and Management System (TMMS) is a growing research area as cities infrastructure continues to evolve. Traffic situation is demanding innovative solutions for effective monitoring and management given the complex nature of the urban scenario. A major focus of this research domain is fine-grained vehicles classification that requires detection and recognition of distinct features of vehicles. Some of these features are semantic based while others are appearance based. One such appearance-based feature of a vehicle is its logo. Logo detection helps with identification of a vehicle’s make during fine-grained classification process. There are various deep learning methods which give good performance for such object detection tasks. However, it is challenging to exploit these methods due to smaller size of logo especially in a surveillance environment. This work firstly presents a deep learning-based approach for detection of vehicles’ logos in camera video feeds. Due to small size of logos, a unique pipeline using three different deep learning models is designed. Firstly, a modified Improved Warped Planar Object Detection Network (IWPOD-NET) selects a Region of Interest (ROI) and adjusts the orientation of vehicle logo. Then YOLO (You Only Look Once) v5 is used to detect the logo part in the selected ROI and finally, EfficientNet is used to further classify logo into different classes. This pipeline is tested on four surveillance environments namely toll control, law enforcement, dashcam, and parking lot access control. Comparative analysis shows accuracy improvement with this proposed approach in each testing case. A pose variance analysis is also performed to determine the orientation limits to which this approach can work. Secondly, a custom dataset, VL-10 (Vehicle Logos) is presented which provided further insights into the challenges w.r.t local environment settings. The whole approach improved the overall performance of the logo detection and recognition system.