基于微控制器改进SqueezeNet交通标志识别的研究

针对目前交通标志的识别都是基于操作系统之上,无法做到自主可控、稳定可靠的问题,故提出一种基于微控制器卷积神经网络交通标志识别。考虑到微控制器内存及计算速度,研究采用改进SqueezeNet网络模型结构,将PC训练机训练好的各种交通标志权值矩阵文件缩小了50倍,移植到前端Cortex-M核系列开发板上；利用内嵌的CMSIS-NN网络函数库搭建与训练机相同的网络模型结构实现对标志的快速识别。实验结果表明,基于微控制器改进SqueezeNet交通标志识别方法平均识别率达到97.4%以上,识别速度得到了有效的提高, 同时为智慧交通的标志识别提供了一种可选择方案。

Research on surveillance video vehicle type recognition based on Cortex-M Convolutional Neural Network

At present, the recognition of traffic signs is based on the operating system, which cannot achieve autonomous control, stable and reliable. Based on this, a method of traffic sign recognition based on Microcontroller convolutional neural network was proposed. Considering the memory and calculation speed of the microcontroller, the research uses the improved SqueezeNet network model structure to reduce the weight of the traffic sign matrix files trained by the PC training machine by 50 times, and transplanted to the front-end Cortex-M core series development board; The embedded CMSIS-NN network function library is used to build the same network model structure as the training machine to realize fast recognition of the sign. The experimental results show that the average recognition rate of SqueezeNet traffic sign recognition method based on Microcontroller is over 97.4%, and the recognition speed is effectively improved. At the same time, it provides an alternative scheme for intelligent traffic sign recognition.