Real-time vehicle detection and classification using deep-transfer learning

Nowadays, vision systems are increasingly present in the road context, particularly in surveillance cameras. These systems help ensure safety, monitor traffic, detect congestion, and facilitate mobility. To optimize the performance of these vision systems, it is necessary to have a reliable detection and classification system. Following the great success of convolutional neural networks, we proposed a machine learning-based road images classification approach, using transfer learning, to address the problem of vehicle detection and classification and provide real-time traffic statistics. In this work, the detection and classification of vehicles have been performed using various deep neural network algorithms, including Faster R-CNN, ResNet, YOLO, SSD, and VGG16. The method can differentiate between five classes: bicycle, motorbike, car, bus, and truck.The deep learning method’s training and testing phases were carried out using a database composed of images captured from videos recorded by ADM Morocco, as well as images from GitHub, Kaggle, and the database called CEREMA. Obtained results reached a classification accuracy of 83. 54% using YOLO. For models based on Faster R-CNN for detection, classification with ResNet reached 90. 15% accuracy, while using VGG resulted in 90. 14%. Similarly, SSD models combined with VGG and ResNet both obtained an accuracy of 87. 73%. The best performance was achieved by combining Faster R-CNN and SSD for detection, followed by ResNet for classification, which yielded an accuracy of 90. 31%.