Abstract
A moveable concrete barrier on the Auckland Harbour Bridge facilitates traffic flow control and optimisation. The concrete barrier's block segments are inter-connected with metal pins, which sometimes can pop out of their safe position. This thesis aims to use deep learning to assist visual metal pin inspection to improve traffic safety. The thesis proposes real-time pin status detection and alerting solutions using various types of video sources. The first part of the proposed network detects and classifies the unsafe pins. The second part actively tracks and alerts the user of unsafe pin status. Preliminary experiments on a small dataset indicated that we could detect unsafe pin status with high precision and recall.
The novel contributions presented in the thesis include: (1) A universal system globally applicable to similar traffic flow regulation and safety contexts with minimal modifications. (2) A novel technique for obtaining synthetic frames to produce different degrees of unsafe pin images obtained from the original video frames. Collectively, synthetic minority-class data boosting, adaptive, incremental, and transfer learning utilising pre-trained neural net-works allow a robust approach to data analysis and modelling on initially small and unbalanced datasets for circumstances where the expected size of the dataset may or may not become available within the expected timeframes (such as during the pandemic lockdowns and added safety requirements). From the presented proof-of-concept, future work is intended to include collaborative user-centred design, where models, software upgrades and analytical platform upgrades will be under the oversight of New Zealand NZ Transport Agency and Auckland System Management.
