Artificial Intelligence for Construction safety

1. Introduction:

Construction workers are killed on the job five times more often than other Workers. According to OSHA, the leading causes of private sector deaths (excluding highway collisions) in the construction industry were falls, followed by struck by an object, electrocution, and caught-in/between.

This issue of security in the construction enterprise is now searching towards the science of Artificial Intelligence (AI) to locate solutions. Artificial brain permits equipment like picture recognition, to acquire web site associated statistics and type through the photographs, videos, and different sources to label feasible risks. With a growing quantity of data, the algorithms can make bigger the accuracy of predictions and minimize the risks. Contractors are increasingly more opting for AI-driven protection setups. The ability of this to realize threats and improve itself over time lends the indispensable envelope of protection that is applicable in the construction industry.

According to a study, a enterprise that takes proactive measures can curb incidents by at least 85%. On the different hand, the businesses that rely on past episodes of injuries or accidents have a a good deal lesser chance of curbing in addition safety issues.

Computer vision (CV) is a field of artificial intelligence that trains computers to interpret and understand the visual world. Using digital images from cameras and videos and deep learning models, machines can accurately identify and classify objects — and then react to what they “see.”

Computer vision is changing the way businesses operate. From retail to warehousing, from manufacturing to urban planning, the technology allows to measure and analyze real life objects and create a quantifiable data level on top of them. 

Computer vision utilizes video or image data to ‘count the uncountable’. The software can analyze video footage from video surveillance cameras which are already present at most modern workplaces. It is simple to deploy, requires little to none additional hardware investment. The advances in machine learning make computer vision applications more accurate and dependable. 

Therefore video analysis is making inroads into construction sites. These tools can increase productivity, improve safety, reduce costs and lead to better informed planning decisions. 

1.1 Applicability of AI and Computer vision in Construction Safety:

1. Monitoring PPE compliance

Everyone knows that even the best safety protocols ‘on paper’ are hard to enforce in real life. Computer vision can identify if hardhats, high visibility jackets, gloves, masks, footwear and other protective equipment is worn by employees entering company premises or certain areas where such equipment is required. It can automatically notify the site security officer if a PPE compliance breach was detected. The video footage can be also stored for later incident analysis. The technology does not take coffee breaks, never sleeps and provides constant site monitoring.

2. Tracking material quantities

Computer vision can track the quantities of building materials stored on the construction site. The technology can provide alerts in advance when quantities are running low. 

3. Preventing unauthorized access

Computer vision solutions are already utilized in many worksites for access control. By identifying license plates, it can prevent unauthorized vehicle access. Taking things further, technology also utilizes face recognition and can automatically identify unauthorized persons. 

4. People and equipment Tracking

Computer vision can track the number of people and equipment passing through a specific scene, and controlling the congestion of construction sites.

1.2 Problem Statement:

In this research, we have utilized the AI and computer vision techniques such as Deep Neural networks to track the labors on-site if they are wearing Helmets or not.

1.3 Solution statement:

The problem has been converted into a two-class (PPE, No PPE) object detection problem and trained a tiny YOLO (a state of the art object detection framework based on Convolution Neural Networks) model to serve the purpose.

A bit of YOLO(You Only Look Once)

YOLO is a real-time object recognition algorithm by “Joseph Redmon et al.” Prior to this object detection was mostly performed by classifiers in two stages:

1.      Divide the image into various regions and find the region of interests (regions where the probability of having an object is high)

2.      Classify these regions using a CNN

The whole process is slow because we need to run prediction through the regions in multiple passes. YOLO proposes an alternate approach, instead of selecting ROI from an image it maps the problem to a regression problem. In YOLO, a single convolutional neural network is applied to the whole image which predicts the bounding boxes and probabilities for each region in a single pass. These predicted bounded boxes are then weighted by the predicted probability. This single-shot prediction makes it much faster than other algorithms like R-CNN, faster R-CNN etc.

YOLO achieves an accuracy of mAP of 48.1 on COCO dataset over a SSD500 model(46.1). Though it achieves higher FPS over two stage detectors, we still did not find it enough for our use case with limited infrastructure(a single GPU). We opted for a lighter version of YOLO called tiny YOLO, which is almost 442% faster than the full version. Although it’s a trade off between accuracy and speed, we were still able to achieve a descent performance.

We will follow the below steps:

4.1 Data Collection:

Collecting quality data is a crucial part of any data science Project, images of people wearing Helmet, people not wearing a Helmet and of both in different Construction sites, backgrounds, angles and scales have been downloaded from Google. Ultimately, more quality data makes model smarter.

4.2 Annotating Images:

We will use “labellmg” tool to create the annotation files in YOLO format as below:

the annotation text file sample(Yolo Format) will be as below :

4.3 Prepare files for training:

We will train our model in colab with a free GPU.

To reduce our computational time will use yolov3-tiny as we have only two classes as below:

We will train until we achieve a decent average loss (0.54), Weights will be saved periodically while training.

We have achieved good results after 2000 iterations:

4.4 Prediction:

We will test our model on different images and videos, and below are our results.

4.4.1 PPE Images Detection:

4.4.2 PPE Recorded Videos Detection:

4.4.3 PPE Real-Time Detection:

4.2 Justification and Conclusion:

·  The Model can be improved by training more images with different backgrounds and different positions.

·  We can build another model to track the full PPE for the Labours.

·  The Applicability of AI and Computer vision in construction can develop safety by detecting unsafe objects and track the people, equipment, and congestion in construction sites. 

5.Reference

The Complete Code on Github:

https://github.com/AhmedsafwatEwida/Real-Time-PPE-Detection

?