| Nowadays,it needs a long-time research to apply autonomous driving technologies in real-life scenario.Hence,assistance driving is the available solution.Recognition algorithms based on visual information is an important technology for assistance driving.The visual signals provide abundant visual information which is useful in recognition process.As the most basic traffic ingredient,traffic signs constrain and guide drivers to drive safely.Therefore,the automatic recognition algorithm for traffic sign becomes one of the most important assistance driving technology which has attracted extensive attention.Most current state-of-art recognition algorithms use the complex deep models which have high requirements for the hardware environment.However,the traffic sign recognition algorithms are mostly distributed on the embedded devices with limited hardware resources in real-life scenario,e.g.mobile phones and vehicle-mounted navigation device.The recognition applications demand highly in real-time and reliability.To handle this scenario,a stable and smaller recognition model is more available.Based on the above considerations,we propose the miniaturization and lightweight methods for the deep model.The miniaturization method is applied on the offline process to construct the simplified deep model.We simplify the complex deep model and design an adhoc training strategy to train the simplified model.The lightweight method compresses the trained model by parameter quantization so that the compressed model can be easily deployed on the embedded devices.Meanwhile the method boosts the real-time performance of the model.The contributions of this paper are:(1)we propose a miniaturization method to construct a simplified convolutional neural network for traffic recognition.This method includes two parts: the model design principle based on multi-scale convolution and concept hierarchy training strategy.(2)A lightweight method based on parameter quantization is proposed.It can further reduce the scale of the trained model and accelerate the speed of the forward inference process.We evaluate the miniaturization method and the lightweight method on the public traffic sign classification dataset and traffic sign detection dataset respectively.The results indicate that our model achieves the comparative performance as the state-of-art recognition model,but our model is smaller and faster. |
