Correctly classifying vehicles as trucks, buses, and passenger vehicles is another application of supervised learning of increased interest to the smart cities industry. While it is difficult to determine the exact features on which a deep learning model relies to determine the likely category with which an image is associated, you can assume that shape is influential. Other features are increasingly important in determining whether a detected shape is a school bus or a metropolitan city bus, such as color or signage. This level of classification specificity might be all that is needed for a high-level traffic analysis in a city. However, other applications of vehicle identification must determine the make, model, color, and year of a vehicle for safety and security. This dramatic increase in the number and similarities of members of each classification level require a more sophisticated data collection, labeling, and training effort.
Convolution neural networks have demonstrated the ability to extract features related to shape, color, and symbols with high accuracy without being programmed with the explicit specification. However, due to the vast number of cars in large markets such as the United States, Europe, and China, successfully training a vehicle identification system requires a carefully curated database of vehicle images together with identifying metadata. This scenario might be simplified by limiting the scope to identification from a single view (head on only), however this approach also severely limits uses.
This use case also generates more research and development of vehicle image databases with corresponding labels for make, model, and year for use with applications that do not rely on vehicle registration records for vehicle identification.