Vehicle Detection and Identification System: Convolutional Neural Network with Self-Attention

Maisha Maliha, Vishal Pramanik

Abstract: The rising number of automobiles on the road worldwide leads to the incidence of accidents and reckless driving. According to the WHO (World Health Organization), road traffic accidents kill around 1.3 million people each year 1. As a result, vehicle detection and number plate recognition systems are essential in terms of investigation. A detecting system is critical in the event of a road accident or a bunch of thugs fleeing with a car. Most of the previous works performed vehicle detection and number plate detection as separate tasks. However, we have taken these two as a single task and created a system to recognize a vehicle and read a number plate. In our work, we have created four models for a real-time vehicle-detecting system where the models classify vehicles into four types: car, bus, ambulance, and truck. We have proposed a novel self-attention-based convolutional neural network framework where the model detects each vehicle by performing several layers of convolution, max-pooling, and self-attention on each pixel with the surrounding pixels of the image. After that we have also compared our model with two different object detection models, VGG16 and YOLOv3. Our result shows that the CNN based self-attention framework outperformed the other models with an overall accuracy of 94.3% as the CNN with self-attention model can accurately decide whether each cell of images related to each other or not.

External IDs:dblp:conf/icccnt/MalihaP23