Deep Learning Model for UAV aided Traffic Analysis and Vehicle Classification

Intelligent transportation systems (ITS) depend upon a range of communication that allows for data exchange between vehicles and traffic information data centers to create applications for providing information like traffic alert in view of lane closure, traffic congestion, track unsafe driving behaviour, optimize routes and emergency management, etc. An important component in this setup are video clips of highways or cross-section areas which needs to be monitored. Traditionally, a closed circuit television (CCTV) is the most often used method of monitoring traffic information, but have limitations like fixed positions, small monitoring region leading to numerous CCTV installations which are costly. A UAV, unlike CCTV, is capable of recording an extensive area of roadways by elevating its altitude or altering its location, as well as travel to a specific spot to monitor unforeseen circumstances such as crashes. Additionally, a UAV observes highways perpendicularly, thus the automobiles in the captured footage do not overlap. This paper explores the issue of vehicle identification on highways using UAV images through deep learning models. Deep learning approaches can automatically learn problem-specific features, eliminating the need for standard feature extraction procedures. Three pre-trained deep learning based model VGG16, ResNet-18 and EfficientNetB3 models have been used for the vehicle classification problem, which is essentially a five class classification problem. The highest accuracy of 97% is reported by the EfficientNet model.