Computer Vision can Help Companies in Manufacturing | CVisionLab

 

How Video Camera System Can Help Companies in Manufacturing

Checking Personal Protection Equipment

Checking Personal Protection Equipment

 

Depending on the industry and working environment, there may be different requirements for workers to follow. For example, wearing personal protective equipment (PPE), HORECA uniforms, technical maintenance, hospital or laboratory clothes, emergency kit equipment or vehicle equipment.

However, it is hard to keep track of every worker and oversee whether they follow the rules. In order to avoid unfortunate consequences and penalties, digitally transform and automate the way you monitor what is happening within the company.

To help employers gain a clear understanding of the internal processes, CVisionLab’s solution utilized computer vision algorithms.

The method is based on deep neural network capabilities for image pattern recognition. This allowed us to spot workers violating health and industrial safety regulations. However, the scope of application extends to many other areas.

It is hard to keep track of every worker and oversee whether they follow the rules.

 

CVisionLab

CVisionLab’s solution understanding of the internal processes

To help employers gain a clear understanding of the internal processes, CVisionLab’s solution utilized computer vision algorithms.

The method is based on deep neural network capabilities for image pattern recognition. This allowed us to spot workers violating health and industrial safety regulations. However, the scope of application extends to many other areas.

Request a commercial offer

 

 

What Surveys Says, How Big is the Issue?

The specialized clothing and other work accessories are put in place to create a barrier against workplace hazards. Unfortunately, a survey by the American Bureau of Labor Statistics revealed that:

 

84%

 

The majority of workers who suffered injuries, such as concussions, skull fractures, and scalp wounds, did not wear head protection.

 

60%

 

A significant portion of eye injuries was associated with not wearing eye protective equipment.


A combination of regular hazard awareness and timely detection of noncompliance can have a great positive impact. Also, it helps companies and workers avoid significant industry penalties. For instance, there are on-thespot fines issued for work health and safety offenses, up to $7,000 for each violation.

 

 

How Computer Vision can Solve PPE Issues or Challenges We Faced

Algorithms for object detection and recognition need to account for different types of activities. As PPE requirements may vary depending on the industry and particular organization, the parameters for image detection should be adjusted accordingly.

CVisionLab on Expro 1
CVisionLab on Expro 2
CVisionLab on Expro 3

 

For example, it can either be one of the following conditions or a combination of them:

  • One piece of required clothing/ accessories being visible
  • An entire set being visible
  • Clothing/accessories worn correctly

In our case, a set of PPE consisted of: a hard hat, protective glasses, a reflective vest.

 

Taking into consideration the three conditions for the image detection above, we needed to include them for every piece of the set:

  • Head protection had to be worn on the head, as opposed to being held in the hands
  • Glasses had to be put on properly, rather than being hooked on a piece of clothing
  • The vest could not be held or thrown over the shoulder with an added parameter for the vest being buttoned-up.

 

 

Data Collection and Systematization

To achieve efficient feature extraction and selection, we compiled a dataset spanning many different object categories. Overall, the number reached about 85,000 images. The images were partially collected from the web. The other part consisted of annotated datasets relying on the collaborative effort of our specialists.

The dataset contained a wide variety of environments and multiple images of different instances of the same class. Alternating shooting conditions and a variety of digital cameras assisted the process further.

 

The ability to work with a broad range of image annotation services, allows us to achieve an effective collection scheme. By extension, our main specialty revolves around quality control of annotations obtained from such services.

 

 

Our Answer to PPE Detection and False Negatives

 

99.6%

accuracy

A major concern we needed to address, for PPE detection, was extreme head positioning. As a person started turning their head sideways, the system stopped recognizing the face within the frame. By using facial detection, we eliminated the problem and achieved 99.6% accuracy.

The results were additionally processed by enforcing temporal consistency. Thus, we minimized instances of misidentified objects caused by unwanted circumstances. For example, the lens could be blocked for a short period of time.

The neural network distinguishes every element of PPE and determines whether the image answers the specified requirements. Every person within the frame is analyzed separately. So, the results are shown in a categorized manner.

 

 

Technical Specifications of the Solution

The performance rate can be regarded as fairly high – up to 20 FPS when using a Full HD camera and a middle-tier graphics card, like GeForce GTX 1050 Ti Mobile. Such cards are widely available in averagely priced laptops. Also, this solution provides simultaneous processing of up to 6 people with no effect on performance. If the number exceeds six it can result in a drop in the framerate.

Model Description Accuracy Recall Precision F-score
MobileNet + softmax, hat, Adam, lr=0.01, cat_crossen, crop by yolo with HD 0.984 0.983 0.997 0.991
MobileNet + softmax, glass, Adam, lr=0.01, cat_crossen, crop by yolo with HD 0.985 0.973 0.982 0.981
MobileNet + softmax, vest, Adam, lr=0.01, cat_crosse, crop by yolo with HD 0.983 0.975 0.989 0.989

Accuracy – is the ratio of number of correct predictions to the total number of input samples.

Precision – is the fraction of relevant instances among the retrieved instances.

Recall – is the fraction of the total amount of relevant instances that were actually retrieved.

F-score – is query classification performance.

 


 

Usability of Detection Results

 

Our detection method has been utilized for an industrial facility with the aim to ensure continuous monitoring. As evidenced in practice, noncompliance with PPE rules can take many forms. For instance, the system detected a number of workers holding hard hats in their hands. This further proved the necessity to analyze the object placement, rather than its pure visibility.

Once the system identifies non-compliance, it can be programmed to perform different actions: from alerting the Security Service to automatically shutting down the operations.

Control over safety violations used to be limited by the human factor. As much as employers would benefit from assigning a safety officer to every worker, this is not feasible. But when you utilize an automated system, with more than 99% accuracy, you receive the benefits of effective and impartial service.

With recent developments in computer vision, you can implement this system into all kinds of production premises. It does not require costly facilities, such as a separate data center. Lastly, there is no need for an IT department to maintain the system since it is fairly simple to manage.

You can implement this system into all kinds of production premises.