A low-speed park unmanned vehicle cruising and emergency braking system based on machine vision

Abstract

The invention discloses a low-speed park unmanned vehicle cruising method based on machine vision. The method comprises the steps that 1) establishing a road semantic segmentation model, and enablingthe semantic segmentation model to adjust a backbone of an ICNet by making a road local data set, training road local data, finely adjusting parameters and a network structure on the basis of the ICNet to reduce the size of a convolution kernel; loading a model, inputting an image to be predicted, and operating model prediction input; and 2) calculating the distance of the vehicle deviating from the road center based on the identification result of the road semantic segmentation model. The method solves the problem of high cost of an automatic driving scheme taking a laser radar as a main sensor, and also solves the problem that lane line perception of traditional computer vision is affected by environmental complexity. By means of the method, due to the fact that the performance of the localized ICNet model is improved, the Yolov3 model and the ICNet model can still operate at the same time, and 20 fps + can still be achieved.

Description

technical field [0001] The invention belongs to the field of deep learning semantic segmentation, and more particularly relates to a low-speed park unmanned vehicle cruise and emergency braking system based on deep learning semantic segmentation. Background technique [0002] Semantic segmentation, which classifies each pixel of the image. The more important semantic segmentation datasets are: VOC2012; and ; MSCOCO; the existing traditional machine learning methods are: pixel-level decision tree classification, refer to TextonForest; and ; Random Forest based classifiers. Then there is the deep learning method. More precisely, convolutional neural networks. The initial popular segmentation method of deep learning is patch classification. The surrounding pixels are extracted pixel by pixel to classify the center pixel. Since the end of the convolutional network at the time used full connected layers, only this pixel-by-pixel segmentation method could be used. In 2014, fr...

AI Technical Summary

Problems solved by technology

[0010] Although the technical solutions of A and B have dealt with the noise of lane line detection to a certain extent, the traditional computer vision solutions still cannot solve the effects of shadows, weather, light intensity, etc.

Images