Multi-scale traffic signal sign identification method based on GMM clustering

Abstract

The invention discloses a multi-scale traffic signal sign identification method based on GMM clustering. The identification method comprises the following steps: obtaining a priori frame size throughGMM clustering, and taking the priori frame size as a parameter of a network to participate in training, training sample images with excessive classes of training data set firstly; inputting an imageto be trained into the neural network, enabling the network to extract feature maps of different levels of the input image, then carrying out upsampling and feature fusion, and finally outputting fiveprediction results of different scales. updating model parameters through iterative training to obtain a transition model; training sample images with few classes according to the method to obtain afinal model; and during identification, inputting a to-be-identified image into the final model to obtain an identification result at a corresponding position of the image. Through GMM clustering, thetraining speed, recognition speed and precision of the network are improved; through multi-scale prediction, the problem that a traffic signal sign is too small and is difficult to detect is solved;through a transfer learning method, the problem of poor recognition effect caused by few data sets is solved.

Description

Technical field [0001] The invention relates to a method for identifying traffic signal signs, in particular to a method for multi-scale identification of traffic signal signs based on GMM clustering. Background technique [0002] Real-time recognition of traffic lights and traffic signs is an important part of autonomous driving and assisted driving technology, and it is also an important part of achieving safe driving. Vehicles need to quickly identify the traffic lights and traffic signs on the road ahead and use them as a basis to make correct driving operations or remind the driver to maintain safe driving. This improves the safety of autonomous driving and reduces the possibility of traffic accidents. [0003] The traditional method is mainly based on the characteristics of the color and shape of traffic signal signs to identify. RGB color segmentation is a recognition method based on color features, but the recognition accuracy of this method is poor. In response to this ...

AI Technical Summary

Problems solved by technology

However, the covariance of each cluster in K-means is zero in all dimensions, so that the clustering result is limited to a circle

However, the two-dimensional data set composed of the length and width of all the calibration frames is irregular in shape. If the circle is used for clustering, the clustering result will have a huge error, which will lead to a serious time-consuming regression process of the prediction frame during the training process. , which affects the training speed and recognition speed; and, the inaccurate clustering results will directly lead to low recognition accuracy of the neural network

[0006] At the same time, most of the existing target detection methods require that the training data set samples are sufficient and even in number. If there are too few data set samples in several categories, the final recognition effect of these categories will be very poor, and false detections and omissions will occur. inspection phenomenon

Images