CN-122023800-A - Mining underground monorail crane track segmentation system and method based on U-Net

Abstract

The invention discloses a mining underground monorail crane track segmentation system and method based on U-Net, and belongs to the technical field of mine intellectualization and computer vision intersection. Aiming at the problems of poor track recognition robustness and low accuracy caused by underground low illumination, high dust and strong light reflection, the invention constructs a segmentation system taking U-Net as a core, wherein an image is acquired through a mining explosion-proof RGB camera; the method comprises the steps of training a model by adopting a data enhancement strategy of brightness degradation, noise injection and dust simulation, realizing pixel level segmentation by utilizing an encoder-decoder structure and jump connection, and optimizing boundary details by a composite loss function. The system only needs to meet the RGB camera of the mining explosion-proof standard, does not need a laser radar, and can run in real time on an embedded platform. Experiments show that the testing set mIoU reaches 87.91%, the typical underground interference performance attenuation is very small, the high robustness, the high precision and the strong engineering practicability are achieved, and a reliable visual perception basis is provided for intelligent navigation of the monorail crane.

Inventors

• WANG HUANYU
• LUO ZEXIN
• YU WENXIN

Assignees

• 湖南科技大学

Dates

Publication Date

20260512

Application Date

20260128

Claims (3)

1. 1. The mining underground monorail crane track segmentation system based on the U-Net is characterized by comprising an image acquisition module, a preprocessing module, a track segmentation core module and a post-processing and decision-making module, wherein the image acquisition module is used for acquiring underground roadway images containing monorail crane tracks in real time, the image acquisition module adopts an RGB industrial camera conforming to mining explosion-proof standards, the preprocessing module is used for carrying out size normalization and pixel value standardization on an original image, the track segmentation core module adopts a U-Net deep neural network model and is used for generating a pixel-level track segmentation probability map, and the post-processing and decision-making module is used for converting the probability map into track area information and outputting the track area information to a control system.
2. 2. The system of claim 1, wherein the U-Net depth neural network model comprises an encoder path, a bottleneck layer and a decoder path, wherein the encoder path is composed of three downsampling blocks, each downsampling block comprises two 3×3 convolution layers and one 2×2 max pooling layer, 112×112×32, 56×56×64 and 28×28×128 feature maps are sequentially output, the bottleneck layer comprises two 3×3 convolution layers, 28×28×256 feature maps are output, the decoder path is composed of three upsampling blocks, each upsampling block is subjected to upsampling through 2×2 transpose convolution, then channel splicing with the feature map of the corresponding level of the encoder is performed through jump connection, then two 3×3 convolution layers are subjected to gradual recovery of spatial resolution, and finally 224×224 single channel segmentation probability maps are output through the 1×1 convolution layers and Sigmoid activation functions.
3. 3. The system of claim 1, wherein in the U-Net model training phase, a composite loss function consisting of binary cross entropy loss and position loss is used for optimization, wherein the binary cross entropy loss is used for measuring the overall classification error between the prediction probability and the real label, the position loss is used for improving the overlapping degree of a track area and relieving the imbalance problem of positive and negative samples, the two are linearly weighted and combined with equal weight, and meanwhile, a data enhancement strategy for simulating underground typical interference is introduced in the training process, and comprises the steps of randomly adjusting a brightness channel in an HSV color space to simulate brightness degradation, superposing Gaussian noise and spiced salt noise to simulate sensor noise and particulate matter interference, superposing semitransparent dark brown or gray textures in a local image area and applying Gaussian blur to simulate dust shielding effect.

Description

Mining underground monorail crane track segmentation system and method based on U-Net Technical Field The invention relates to a mining underground monorail crane track segmentation method based on a U-Net deep learning model, and belongs to the technical field of mine intellectualization and computer vision intersection. Background In underground coal mining operations, suspended monorail cranes are key devices for personnel and material transportation. However, the operation is highly dependent on manual operation, and in complex underground environments with low illumination, high dust, moisture and a large amount of metal structure interference, drivers are extremely prone to collision, derailment and other safety accidents caused by visual fatigue or poor visibility. The method realizes the automation or unmanned operation of the monorail crane, and has the primary premise of constructing a reliable visual perception system and being capable of accurately identifying and positioning the track. Traditional computer vision methods, such as edge detection methods based on Canny operators, sobel operators, etc., or straight line/curve fitting in combination with Hough transform, rely heavily on clear, continuous edge features. In an underground actual scene, the track is often blurred or even missing in edge information due to rust, greasy dirt, water stain or partial shielding by dust, so that the traditional methods fail and have poor robustness. In recent years, a deep learning driven semantic segmentation technology provides a new idea for solving the problem. Among them, U-Net has achieved great success in the field of medical image segmentation and the like by virtue of its unique encoder-Decoder (Encoder-Decoder) symmetrical structure and Skip Connections (Skip Connections) mechanism. The structure can effectively fuse global context information and local space details of the image, and is very suitable for processing tasks needing accurate positioning. However, the direct application of U-Net to non-uniformly illuminated and dynamically-interfering downhole industrial scenarios still face significant challenges. The prior art lacks a special segmentation system which is specially designed for the characteristics of the monorail hanging rail of the mine and can effectively resist the typical underground interference (such as dust, noise and extreme illumination). Disclosure of Invention In order to solve the problems, the invention provides a mining underground monorail crane track segmentation system based on U-Net, which has clear structure and convenient deployment, and provides a mining underground monorail crane track segmentation method with high algorithm efficiency, strong anti-interference capability and accurate segmentation boundary. The technical scheme of the invention for solving the problems is that the mining underground monorail crane track segmentation system based on the U-Net comprises an image acquisition module, a preprocessing module, a track segmentation core module and a post-processing and decision-making module, wherein the image acquisition module is used for acquiring an underground roadway image containing monorail crane tracks in real time, the preprocessing module is used for carrying out standardized processing on an original image, the track segmentation core module adopts a U-Net deep neural network model and is used for generating a pixel-level track segmentation probability map, and the post-processing and decision-making module is used for converting the probability map into available track area information and outputting the available track area information to a control system. A mining underground monorail crane track segmentation method based on U-Net comprises the following steps: firstly, acquiring RGB images containing monorail crane tracks in underground roadways through an industrial camera arranged at the front end of a monorail crane, and carrying out pixel-level artificial labeling on the acquired images to construct a training data set containing track mask labels; Step two, data preprocessing and enhancement, namely performing size normalization and pixel value standardization on an original image, and introducing a data enhancement strategy for simulating underground typical interference in a training stage, wherein the data enhancement strategy comprises brightness degradation, noise injection and dust shielding simulation; Building and training a U-Net neural network model with an encoder-decoder symmetrical structure and a jump connection mechanism, and optimally training by adopting a composite loss function formed by binary cross entropy loss and Dice loss until the model converges; converting the trained U-Net model into a lightweight format, deploying the lightweight format to a vehicle-mounted embedded computing platform, receiving a real-time image stream when running, and performing forward reasoning to generate a track segmentation resul