One step implementation of yolox helmet detection (combined with oak intelligent depth camera)

This article was first published in oakchina.cn

Safety production is the eternal theme of social development , However, the traditional safety production supervision methods exist Large monitoring blind area 、 Low supervision efficiency 、 Management difficulties of outsourcing personnel, etc shortcoming . And the development of deep learning , Bring new solutions to safety production .

Preface

In accordance with the relevant provisions , Wearing safety helmet correctly is of great significance to safety production , But the traditional helmet wearing inspection needs a lot of manpower and material resources , There are many inconveniences in the actual inspection process . Use the deep learning network model to detect the wearing of safety helmet , It can make up for the shortcomings of traditional methods . Specific application scenarios include ：

• Safe production in high-risk work areas ： Deploy helmet wearing detection system in high-risk work areas , Monitor the wearing of safety helmets of on-the-job personnel in real time through the monitoring screen , If the safety helmet is not worn as required , be Output alarm information , Automatically save the screenshot of the staff not wearing the safety helmet as a certificate , And inform the background monitoring personnel . The application can be through the video monitoring system , Real time monitoring the wearing of safety helmets of staff in high-risk work areas , It can not only ensure the smooth progress of safety production , It can also reduce the difficulty of supervision and management cost .
• Smart construction site ： Integrate the helmet detection system with the detection of falling objects 、 Combined with the staff fatigue detection system , Build a smart construction site , Give early warning of potential hazards in time , Inform relevant personnel to take measures , So as to stop personal injury accidents in time , Ensure safe production .

analysis

How to define “ Safety helmet wearing test ” This question ？ Normally, it is necessary to judge whether there is safety helmet in the pedestrian head area , For example, some methods are to use the target detection model first, such as SSD、Yolo Series and improvement series （ All kinds of accelerators and mobile networks ） Detect pedestrian areas , Then we design a small classification network to determine whether there is a helmet in the area . The advantage of this approach is that it is relatively simple ：

• in the first place , Pedestrian detection is a more widely used method , Data sets 、 There are many algorithms ;
• both , A lot of data is in the form of surveillance video , When marking, mark the area with coarser granularity , Then make a classification .

But the disadvantages of this approach are also obvious ：

First of all , This is not an end-to-end prediction process , Obviously, it is necessary to detect the travelers first , Then use the classified network to judge whether there is a helmet ;
second , The difference between target detection and classification , If we use this coarse-grained marking method , That is, mark out the larger area containing the helmet , Not the area that contains the helmet , The effect of classification is not so good ;
Third , Strictly speaking , What users want more is , Locate the target position of wearing safety helmet , In many cases, helmets are in pedestrian areas , But it's not out of “ Wear ” state .

Based on the above analysis , The method used here is to pay attention to safety helmets “ Wear ” state , So at the business level , The data needed is “ Wear ” Helmets and “ Wear ” Pictures of helmets . As shown in the figure below ：

One of them is red bounding box It's a target without a helmet , Green bounding
box It's the goal of wearing a helmet , This can be more accurate 、 More directly determine whether to wear the State , However, we need to cover more positive and negative target data of the scene .

Realization

Based on the previous analysis , The task here can be seen as a routine case of target detection , The algorithm of target detection develops very fast , A single-stage, multi-stage , Various anchor And anchor-free Of ,GIoU,soft-NMS, Multi scale prediction of various pyramids and so on , But a lot of sota It's very difficult to apply the method of , After all, speed 、 Power waste 、 Cost and other factors should be taken into account , This article only uses YOLOX-nano Make an example , The renderings are as follows ：

use OAK Do a helmet test

Source code ：helmet_yolox
Hardware ：OAK-D

---

OAK China | track AI New developments in technology and products
official account | OAK Visual artificial intelligence development
spot 「 here 」 Add wechat friends （ Remarks ）
stamp 「+ Focus on 」 Get the latest information