Abstract:
Understanding the precise movements of different commodities on South African roads can help in not only describing the logistics sector more accurately, but also in the planning of road infrastructure maintenance and investment. Truck combinations can be classified into several classes broadly associated with different commodity groups, including tautliners, tankers, flatbeds (general freight) and flatbed (containerised freight). Current truck classification systems in South Africa can classify trucks by number of axles and vehicle mass but are unable to determine the combination type and hence commodity group. Video data allows for truck combinations to be classified in more detail using image-based classifiers. The latest developments in deep learning algorithms have made it possible for accurate classification of vehicle types using camera data. A CCTV camera feed of a section of the N3 was provided by the South African National Roads Agency Limited (SANRAL) and was used as a case study to develop a proof-of-concept classifier for tautliner and tanker truck combinations, using a transfer learning approach and the pre-trained ResNet50 classifier. The results indicate good accuracy based on relatively small datasets. Future work will focus on further optimisation and investigating the training dataset requirements in more detail.
