Traffic target detection and identificationrecognition method self-adaptive to scene change

Abstract

A traffic target detection and identification method adaptive to scene change comprises the following steps: 1) constructing an input data set exclusive to road traffic, the input data comprising images of a color traffic scene, label files corresponding to the images, and indexes of pictures for training and testing; 1.1) for the image, calculating the contrast of the image, performing de-lighting enhancement processing on the image with low contrast, and highlighting the edge and contour information of the target; 1.2) for the label file corresponding to the image, determining the size of a target contained in the label file in the training data and the distribution rule of the target by using a statistical and dimensional clustering method; 2) enabling the enhanced image to enter an improved neural network, and generating feature maps with different scales; 3) on the generated feature map, modifying the proportion and the size of a preselected box in the neural network according to the statistical and dimensional clustering results in the step 1.2) to generate the preselected box; and 4) inhibiting and screening an output result through a non-maximum value, subtracting a prediction result from a true value in the tag file to calculate a loss function, carrying out back propagation on an error, and updating a neural network parameter.

Description

[0001] Technical field: [0002] The invention belongs to the fields of computer vision and target detection, and more particularly relates to a traffic target recognition method for solving adaptive scene changes. [0003] technical background: [0004] With the continuous advancement of technology and the enhancement of computer computing power, more and more object detection algorithms based on deep learning are applied to vehicle detection. In the face of daily poor light conditions such as rainy and dim, existing algorithms often have problems of missed detection and false detection. At present, there are two main ideas to solve this problem. One is to supplement the training samples so that the model has better generalization ability to adapt to scenes under different lighting conditions; the other is to increase the complexity of the training network so that It extracts deep features from training samples. These two ideas can solve the problem of low model accuracy cau...

AI Technical Summary

Problems solved by technology

These two ideas can solve the problem of low model accuracy caused by poor lighting conditions to a certain extent, but they cannot solve the problem fundamentally.

Images