Vehicle Type Recognition using Capsule Network

Vehicle type recognition plays an increasingly important role in the problem of dealing with traffic safety. In order to address this problem, a novel vehicle type recognition model based on capsule network is proposed, which dynamic routing algorithm of the capsule network is redesigned to save training time and to accelerate convergence rate. To compare with the original model, we conduct experiments on the dataset including five different vehicle types from different angles and lighting conditions. It is proved that the proposed model costs only half time of original model while maintaining a high recognition rate, and the reconstruction from the output of capsule network keeps important details of vehicle types.