Single class detection-based deep learning approach for identification of road safety attributes

Automatic detection of road safety attributes is an important step in designing a reliable road safety system. Due to the outstanding performance over the handcraft feature extraction-based methods for detecting objects, deep learning can be used to develop a robust road safety system. However, there are many challenges in using deep learning models. Firstly, they require a large dataset for training. Secondly, a class imbalance is a common problem in deep learning models. Finally, when a new attribute is introduced to a deep learning model, the whole model must be re-trained using all training samples which requires a lot of time. In order to solve some of these problems, we propose a novel single class detection-based deep learning approach for the identification of safety attributes in roadside video data. The approach is based on fusion of multiple fully convolutional network (FCN) models. Each model is trained to detect a single attribute/class using two classes (single attribute vs all other attributes) datasets. The proposed approach was evaluated on data provided by the Department of Transport and Main Roads (DTMR). The proposed approach achieved high accuracy and a new attribute can be added to the system without retraining the whole system.