CN-116935184-B - Irregular traffic road target detection method and system

CN116935184BCN 116935184 BCN116935184 BCN 116935184BCN-116935184-B

Abstract

The invention relates to the technical field of target detection, in particular to a method and a system for detecting an irregular traffic road target, wherein the method comprises the steps of collecting and manufacturing a data set, and dividing the data set into a training set, a verification set and a test set; the method comprises the steps of constructing an irregular traffic road target detection network model, preprocessing data of a training set to obtain corner information and offset information, adding the corner information and the offset information to a data label, importing the corner information and the offset information into the irregular traffic road target detection network model for training, substituting a test set into the irregular traffic road target detection network model for prediction, calculating model accuracy, inputting the irregular traffic road target detection network model according to an obtained image, judging the position of a road, and prompting a visually impaired user to go out. The invention realizes the target detection of the high-precision and lightweight road contour, and simultaneously adopts the road position judging method to assist visually impaired people to realize safe traveling on traffic roads.

Inventors

BU JIAJUN
ZHANG YUEQING
YU ZHI
ZHOU CHENG
GU JINGJUN
WANG WEI

Assignees

浙江大学

Dates

Publication Date: 20260512
Application Date: 20230719

Claims (8)

1. A method for detecting an irregular traffic road target, comprising: S1, collecting and manufacturing a data set, and dividing the data set into a training set, a verification set and a test set; S2, constructing an irregular traffic road target detection network model; s3, preprocessing the data of the training set to obtain corner information and offset information, and adding the corner information and the offset information to the data label to be imported into an irregular traffic road target detection network model for training; s4, substituting the test set into the irregular traffic road target detection network model to predict, and calculating the accuracy of the irregular traffic road target detection network model; s5, inputting an irregular traffic road target detection network model according to the acquired image, judging the position of the road, and prompting the visually impaired user to go out; The irregular traffic road target detection network model comprises a main network, a neck network and a head network; the irregular traffic road target detection network model comprises: Constructing a backbone network, wherein the backbone network is used for extracting core characteristics; Constructing a neck network for processing and delivering image features; A header network is constructed for predicting image features.
2. The method for detecting irregular traffic road targets according to claim 1, wherein in S1, the collecting and creating the data set includes that the real-time equipment of the visually impaired cannot support the collection and calculation of the high-resolution image, after finishing the data arrangement, all the data are processed in a unified resolution, then the target corner points and the outline border are marked in a multi-point marking format, and the marked file is output in json format.
3. The irregular traffic road object detection method according to claim 1, wherein the backbone network comprises FASTERNET BLOCK, an embedding module and a merging module, each FASTERNET BLOCK is preceded by an embedding module or a merging module, the embedding module and the merging module each encapsulate a two-dimensional standard convolution layer, a BN layer and a ReLU activation function layer, the convolution kernel size of the convolution layer in the merging module is 2 x 2, the step size is 2, the convolution kernel size of the convolution layer in the embedding module is 4 x 4, and the step size is 4.
4. The irregular traffic road object detection method of claim 3 wherein the characteristics output by the last FASTERNET BLOCK module in the backbone network are input directly into the neck network.
5. The irregular traffic road object detection method of claim 3 wherein the neck network includes a CBS module, SPPCSPC module and ELAN module; the CBS module encapsulates a two-dimensional standard convolution layer, a BN (Batch Normalization) layer and a SiLU activation function layer; The SPPCSPC module comprises two branches, one branch is of an SPP structure, receptive fields of different sizes are obtained through four maximum value pooling submodules of different scales, the other branch only passes through one CBS module of 1 multiplied by 1, and finally the two branches are subjected to feature fusion; The ELAN module comprises two branches and four characteristic paths, wherein the first branch directly passes through a 1X 1 CBS module, the second branch comprises a 1X 1 CBS module and four 3X 3 CBS modules, the three paths respectively extract output from the first, third and fifth CBS modules and are fused with output of the first branch, the size of the characteristic diagram of the output is consistent with that of the input, and the number of channels is halved.
6. The irregular traffic road object detection method according to claim 3, wherein four feature layers of different sizes are provided in the neck network.
7. The irregular traffic road target detection method of claim 3, wherein the head network includes corner points and offset heads; the corner head comprises a 3 multiplied by 3 CBS module, a ReLU activation function layer and a 1 multiplied by 1 CBS module, the corner head is used for predicting Gaussian heatmap of the corner, and the output channel number is 1; the offset head comprises a 3×3 CBS module, a ReLU activation function layer and a 1×1 CBS module, and is used for predicting the offset from the corner point to the abscissa of the traffic road target center point, and the number of output channels is 2, which respectively represents the offset in the horizontal and vertical directions.
8. An irregular traffic road object detection system, characterized in that it is applied to the irregular traffic road object detection method as set forth in claim 1, the three-dimensional object detection system comprising: the acquisition module is used for acquiring and manufacturing a data set and dividing the data set into a training set, a verification set and a test set; the modeling module is used for constructing an irregular traffic road target detection network model; the preprocessing module is used for preprocessing the data of the training set to obtain corner information and offset information; the training module is used for adding the corner information and the offset information to the data label and importing the corner information and the offset information into the irregular traffic road target detection network model for training; The test module is used for substituting the test set into the irregular traffic road target detection network model to predict, and calculating the accuracy of the irregular traffic road target detection network model; The road position judging module is used for inputting an irregular traffic road target detection network model according to the acquired image, judging the position of the road, and prompting the visually impaired user to go out.

Description

Irregular traffic road target detection method and system Technical Field The invention relates to the technical field of target detection, in particular to a method and a system for detecting an irregular traffic road target. Background Many people with vision impairment have inconvenience in life, and one of them is going out. The main factor of inconvenient travel of visually impaired people is the inability to observe road conditions and to find roads to go to the destination. Because the traffic road target seen under the horizontal view angle shows a perspective effect of near, far and small, and the uncertainty exists in the standing position of visually impaired people and the direction of the traffic road, the traffic road often props against the frame or bottom angle of the frame in the view field picture, and an irregular polygon is shown. The existing target detection comprises a traditional target detection method regression rectangular detection frame, a detection method of a pixel level, an example segmentation and the like by a semantic segmentation technology. The conventional target detection method is adopted to return to the rectangular detection frame, although the detection with certain precision can be realized, the outer contour of the traffic road cannot be framed, the position and the direction of the road cannot be distinguished, the semantic segmentation technology is a pixel-level detection method, the contour of the traffic road can be marked, but the accuracy of the semantic segmentation on the target boundary is lower, different target individuals which are shielded or closely arranged in the same category cannot be distinguished, and the example segmentation can mark the contour of the road and the different target individuals, but the light weight degree and the real-time performance do not meet the requirement of moving trip blind guiding of visually impaired people. It is therefore necessary to provide a new irregular traffic road object detection method and system. Disclosure of Invention Based on the above-mentioned problems existing in the prior art, an object of the present invention is to provide an irregular traffic road target detection method, which realizes high-precision and lightweight target detection of the outer contour of the road, and meanwhile, adopts a road position judgment method to assist visually impaired people in realizing safe traveling on the traffic road. In order to achieve the purpose, the technical scheme adopted by the invention is that the irregular traffic road target detection method comprises the following steps: S1, collecting and manufacturing a data set, and dividing the data set into a training set, a verification set and a test set; S2, constructing an irregular traffic road target detection network model; s3, preprocessing the data of the training set to obtain corner information and offset information, and adding the corner information and the offset information to the data label to be imported into an irregular traffic road target detection network model for training; s4, substituting the test set into the irregular traffic road target detection network model to predict, and calculating the accuracy of the irregular traffic road target detection network model; S5, inputting an irregular traffic road target detection network model according to the acquired image, judging the position of the road, and prompting the visually impaired user to go out. Further, in S1, the step of collecting and manufacturing the data set comprises the steps that as the mobile real-time equipment of the visually impaired cannot support the collection and calculation of the high-resolution image, after finishing data arrangement, all data are processed in a unified resolution mode, then the target corner points and the outline border are marked in a multi-point marking format, and the marked files are output in a json format. Further, in S2, the irregular traffic road target detection network model comprises a backbone network, a neck network and a head network. Further, in S2, the irregular traffic road target detection network model includes: S2-1, constructing a backbone network, wherein the backbone network is used for extracting core characteristics; s2-2, constructing a neck network, wherein the neck network is used for processing and transmitting image characteristics; S2-3, constructing a head network, wherein the head network is used for predicting image characteristics. Further, the backbone network includes FASTERNET BLOCK, an embedding module and a merging module, each of which is connected with an embedding module or the merging module before FASTERNET BLOCK, the embedding module and the merging module each encapsulate a two-dimensional standard convolution layer, a BN layer and a ReLU activation function layer, the convolution kernel size of the convolution layer in the merging module is 2×2, the step size is 2, the convolution kernel size