Automobile identification method based on Centernet

Abstract

The invention discloses an automobile identification method based on the Centernet. The automobile identification method comprises the following steps: S1, acquiring and collecting original image data; s2, classifying the original images into a training set, a verification set and a test set; respectively carrying out enhancement processing on the training set, the verification set and the test set to obtain an enhanced training set, an enhanced verification set and an enhanced test set; s3, building a network structure based on the Centernet; s4, standardizing the Center network structure byadopting Group Normalization, taking a Radam optimizer as an optimization method of network training, and carrying out iterative training to obtain a trained Centernet network structure; and S5, verifying and testing the trained Centernet network structure by using the enhanced verification set and the enhanced test set. According to the method, on the basis of vehicle types and position features,the features of the bottom layer in the image are extracted through a novel convolutional network method, and therefore the more detailed features of the vehicle track are obtained at the spatial position.

Description

technical field [0001] The invention belongs to image recognition technology in automatic driving, in particular to a Centernet-based car recognition method. Background technique [0002] The existing automatic driving technology is mainly used in automobiles, trains, etc. as assisted driving, all of which adopt methods such as image recognition and decision-making planning based on deep learning. Related technologies are mainly divided into three categories: (1) SLAM method based on laser sensor, which belongs to the perception method in perception-planning-control; (2) visual SLAM method based on 2D camera sensor; (3) fusion based on multi-sensor method. Different kinds of sensors determine the type of related methods. [0003] Their respective advantages and disadvantages: Laser sensors can more accurately capture distant object information, but the cost is relatively high, and related methods are not as mature as 2D. The 2D camera needs to obtain the obstacle distance...

AI Technical Summary

Problems solved by technology

[0004] The technical problem to be solved by the present invention is that the existing image recognition methods only have mature applications in the recognition of features such as vehicle classification, license plate recognition, and car location, and cannot fully mine the information of 2D images. The purpose is to provide a Centernet-based car identification method to solve the above problems

Images