A Deep Learning Technique for the Analysis of Road Safety Attributes

Road safety systems are essential for planning, managing, and improving road infrastructure and reducing the incidence and risk of road accidents. Manual systems used for road safety assessments are inefficient, time-consuming, and prone to error. Some automated systems using sensors, cameras, lidar, and radar to detect nearby obstacles such as vehicles, pedestrians, lane lines, some traffic signs and parking slots, have been introduced to reduce road fatalities by minimising human error. However, the existing available road safety systems are unable to accurately detect all road safety attributes required by the Australian Road Assessment Program (AusRAP), a program launched to establish a safer road system through inspection of high-risk roads, developing star ratings, and safer roads investment plans to mitigate accident risk. Therefore, it is important to explore novel techniques and develop better automated systems which can accurately detect and classify all road safety attributes. This research focuses on the development of a fully automated system which can accurately detect all road safety attributes using roadside video data collected by vehicle mounted cameras. This thesis proposes the following techniques. First, a new single attribute-based segmentation and classification technique to automatically identify AusRAP attributes is proposed. Specific focus was deep learning techniques, as recent literature suggests that these techniques can achieve better accuracy than traditional techniques. Various deep learning architectures and learning methods were investigated to develop an appropriate deep learning-based technique to detect road safety attributes with high accuracy. Second, the performance of multi attribute-based techniques were further improved by introducing evolutionary classifier-based techniques. Third, the architectures and learning algorithms were optimised by parameter tuning. The suitability of single attribute and multiple attribute-based training methods was investigated. The proposed techniques were evaluated on a real-world dataset, prepared using roadside videos provided by the Department of Transport and Main Roads (DTMR), Queensland, Australia, and the Australian Road Research Board (ARRB). A large number of experiments using the proposed techniques were conducted. The classification accuracy was used as a metric to measure the performance, and to further validate the efficacy different diversity measures such as specificity, sensitivity, and f1-score were used. Appropriate analysis, and a comparison with existing techniques, was conducted and presented. The results and analysis presented in this thesis show that the proposed single-stage and multi-stage deep learning-based techniques not only achieve classification accuracy and mis-classifications better than existing state-of-the-art segmentation and classification techniques, but also provide a platform for future research. It was found through experimentation that by using the single attribute-based training technique proposed, re-training the whole model using all training samples, which requires a lot of time when a new attribute is introduced to a deep learning model, can be avoided. Moreover, through extensive experimentation, it was proved that it is not always required to have a large dataset for training, and data inadequacy and class imbalance problems are no longer issues. Effective solutions were raised to eliminate the requirement to annotate large numbers of samples for each attribute to produce acceptable accuracy for industry while archiving large-scale training speed enhancements. Both single-stage and multi-stage deep learning-based techniques were validated using real-world test data without cropping, and pixel-wise prediction was obtained for each object. The accurate location of the predicted object was known in predictions and hence, bounding box problem was avoided. Through the incorporation of optimisation, optimum parameters suitable for training road safety attributes using real-world datasets were determined. Finally, the optimum parameters proved to be effective not only in terms of classification accuracy, but also confirmed that the proposed optimisation technique converges to the minimum error faster than state-of-the-art techniques.