Large-Truck Safety Warning System Based on Lightweight SSD Model.

Dong Xiao,Qifei He,Hongzong Li,Chenyi Liu

doi:10.1155/2019/2180294

Abstract

Transportation is an important link in the mining process, and large trucks are one of the important tools for mine transportation. Due to their large size and small driving position, large trucks have a blind spot, which is a hidden danger to the safe transportation of mines and has a great impact on production efficiency and economic loss. The traditional large truck safety warning system mainly uses the ultrasonic short-distance ranging method, radar ranging method, GPS (Global Positioning System) technology, and so on. The disadvantage of these methods is that they are affected by the environment and weather, and they cannot display the object status in real time. Therefore, it is becoming increasingly important to realize the large truck safety warning system based on machine vision. Therefore, this paper proposes a lightweight SSD (Single Shot MultiBox Detector) model and an atrous convolution to build a large-truck object recognition model. First, the training images are collected and marked. Then, the object recognition model is established by using the lightweight SSD model. The atrous convolutional layer is introduced to improve small object detection accuracy. In the end, the objectness prior method is used to improve the classification speed. Experimental results show that, compared with the original SSD model, the lightweight SSD model occupies less space and runs faster. The lightweight SSD model with the atrous convolutional layer is more sensitive to small objects and improves detection accuracy. The objectness prior method further improves the identification speed. Compared with the traditional large truck safety warning, the system is not affected by the environment and realizes the visualization of large truck safety warning.

Highlights

With the continuous improvement of mining technology and the continuous increase of mining intensity [1], the large scale of equipment is developing, and the large truck is one of the main tools for mine production and transportation. ere are many characteristics of large mining trucks. e mine transportation road is complex, with more curves and slopes, and the roads are often changed, and the working environment is bad. ere are many kinds of auxiliary vehicles, such as command cars, gunpowder cars, bulldozers, sprinklers, etc
A lightweight SSD model based on atrous convolution is proposed to realize object recognition from the perspective of the truck. e identification of small objects is enhanced while minimizing the volume of the model
Three sets of comparison experiments were carried out based on the selection of the small object feature extraction layer, the expansion rate of the atrous convolution, and the different models. e test results of the lightweight SSD model based on atrous convolution were tested and verified

Summary

Introduction

With the continuous improvement of mining technology and the continuous increase of mining intensity [1], the large scale of equipment is developing, and the large truck is one of the main tools for mine production and transportation. ere are many characteristics of large mining trucks. e mine transportation road is complex, with more curves and slopes, and the roads are often changed, and the working environment is bad. ere are many kinds of auxiliary vehicles, such as command cars, gunpowder cars, bulldozers, sprinklers, etc. A lightweight SSD model based on atrous convolution is proposed to realize object recognition from the perspective of the truck. The atrous convolutional layer is used to add the lower feature layer to multiscale feature fusion, which is helpful for the detection of small objects This model introduces the objectness prior algorithm to improve the model detection speed. E first step is to reduce the resolution with the lower sampling operation of 2, and the convolution operation is performed with a Gaussian kernel of 7 × 7, and the characteristic graph obtained at this time is only one-fourth the size of the original image.

64 Block 1

Experiment

Results and Analysis

Conflicts of Interest