Pedestrian Traffic Lights Classification Using Transfer Learning in Smart City Application

Traffic accidents have become a serious issue in cities. Millions of people die in traffic accidents annually and among them the major cause is the pedestrian jaywalking. To solve this traffic issue and ensure efficient traffic monitoring, we introduced the surveillance system using AI powered UAVs in Internet of flying things based smart city scenario. To accurately classify the pedestrian traffic lights, we use the computer vision technology. We have created our own local dataset containing 809 images where 441 images belong to red signal class while 368 images belong to green signal class. We explore the power of transfer learning based on DNNs to overcome the limitation of dataset for pedestrian traffic lights classification. In this approach, we use the pre-trained MobileNetV2 model and freeze the weights. By leveraging the pre-trained convolutional base, we add our own fully connected layers on top of the model for classification. To handle the problem of limited data, we also perform the data augmentation. The task is formulated as binary classification problem. By using the MobileNetV2 on challenging and very diverse dataset, we achieve the accuracy of 94.92%, 91.84% specificity and 97.10% sensitivity.