Real-Time Traffic Sign Recognition on Sipeed Maix AI Edge Computing

Convolutional neural network (CNN) has become popular and helpful model for classification of objects in the video stream, clustering, pattern recognition and prediction in many disciplines. Considering many systems solving the classification problem, the mobility is often required. This paper proposes an implementation of the Tiny-YOLOv3 (You Only Look Once) using CNN in the Internet of Things (IoT) devices to solve the problem of classification of traffic signs on the EDGE AI platform Sipeed MAIX. The main feature of this module is the availability of 1st RISC-V 64 AI Module include K210 KPU inside, which allows high-performance computing with low power consumption. The proposed method uses K-means algorithm to cluster our training set to find more excellent priori boxes of the targets. Performances of the proposed approach are evaluated on the Belgium Traffic Sign Detection (BTSD) dataset. Extensive experiments demonstrate that the proposed method achieves an excellent trade-off between speed and accuracy. The system executes the detection/recognition process in approximately 112 ms for each image in (KPU) and 9 FPS in video stream.