PT module -A Traffic Signal Classification Model Based on Convolutional Neural Networks and Random Forests
Keyu Ren, Heqing Peng, Junwei Wu, Yao Shengtao, Jinfeng Li, Pingyu Li · Applied and Computational Engineering · 2023
Autonomous driving and image recognition are the hot directions of Internet development nowadays. In self-driving cars it is necessary to capture traffic signs in front of the vehicle by cameras. To ensure that the information of image recognition is correct, a set of image classification models with high accuracy should be used to classify the recognized objects in order to determine the next instruction of vehicle operation. There were many achievements in the research work of traffic sign recognition, but there are still some shortcomings. By combining CNN and random forests with PT module (a module that can improve the accuracy of feature extraction), we finally came up with a classification model that can efficiently place the received traffic sign images into a specified category, which we obtained 97% accuracy on the GTSRB dataset, which is much more accurate than traditional neural network methods or regression methods. We have also evaluated it on other datasets and the results obtained are more promising.