Real-Time Traffic Sign Recognition Based On Cascade Region Proposal And Tiny CNN Group

The level of intelligence and electronics in automobiles continues to improve,and Advanced Driving Assistance System(ADAS)has been widely deployed in mid-to-high-end automotive products.Traffic sign recognition is an important part of ADAS’s environmental perception,which can help drivers pay attention to road condition information in time,remind and assist drivers to make targeted adjustments to driving behaviors,and improve vehicle driving safety.Due to the limitations of power consumption,size and cost,the computing power of the vehicle-mounted computing platform is still at a low level.For traffic sign recognition tasks,some general object detection algorithms cannot achieve satisfactory real-time performance and accuracy on the vehicle-mounted computing platform.Therefore,it is necessary to design a real-time traffic sign recognition method with low computing power requirements and high accuracy.To achieve this,we must comprehensively consider the characteristics of the traffic sign recognition task and the computing power of the vehicle-mounted computing platform.The thesis combines statistical learning and deep learning,and proposes a real-time traffic sign recognition method with small model size,good real-time performance and high precision.The main work content is as follows:(1)Analyze the visual features and image operators of traffic sign images,and propose a Cascade Region Proposal(CRP)for traffic sign detection based on MB-LBP operator and Gentle-Adaboost algorithm.The algorithm is a cascade of multiple Gentle-Adaboost strong classifiers,and measures such as shape mutual exclusion and hierarchical sliding windows are designed to improve performance.CRP can achieve accurate traffic sign detection with excellent real-time performance.(2)A tiny CNN group(TCG)with small parameter amount is designed to realize the classification of traffic signs.The CNN model in TCG has 4 convolutional layers,2 maximum pooling layers,and 2 Dens layers.It adopts multi-class cross-entropy loss function,and the size of a single model is only about 4.3MB.At the same time,non-maximum suppression NMS is used to eliminate redundant candidate boundingboxes.For video data,dynamic recognition compensation is adopted to further improve the real-time performance of the method.The thesis evaluates the proposed real-time traffic sign recognition method on the public dataset TT100 K and an actual road video,and designs an ablation experiment to test the effectiveness of the components.Experiments show that the proposed method can effectively recognize 151 types of traffic signs in the dataset.It achieves 87.45% precision,88.04% recall and 90.57% alarm rate on the 48 types of traffic signs that account for the largest proportion,which is better than some up-to date object detection methods such as SSD,YOLOv4 and EfficienDet.In video test,the proposed method can achieve a frame rate of 42 fps in a test environment that does not include GPU,which shows that the proposed method has excellent performance under low computing power conditions.