Modified-vehicle detection and localization model for autonomous vehicle traffic system

Abstract

The modification of vehicles for financial gain is an evolving tendency observed in India. Recognizing and detecting of these modified illicit cars is an important but critical task in autonomous vehicles (AV). It is always possible for a cyclist or pedestrian to traverse obstacles or other fixed objects that appear in front of any moving vehicle. Vehicles that are autonomous or self-driving require a different system to quickly identify both stationary and moving objects. A deep learning model named you only look once version 5 (YOLOv5)-convolutional block attention module (CBAM) is proposed here for the Indian traffic system which is based on YOLOv5m. The proposed algorithm, YOLOv5-CBAM, has three major components. The first module, the backbone module is employed for feature extraction. The second module is to detect static as well as dynamic objects at the same time and the third CBAM module is adopted in the backbone and neck part, which mainly focuses on the more prominent features. Two cross stage partial (CSP) modules were used after every convolutional layer resulting in an additional head to the proposed model. Four head modules equipped with anchor boxes performed the final detection. For the present dataset, the proposed model showed 98.2% mean average precision (mAP), 98.4% precision, and 94.8% recall as compared to the original YOLOv5m.