Safety in construction is a vital matter that impacts the lives and welfare of both employees and the public. Although it is a crucial part of construction safety management, hazard detection is frequently hampered by environmental constraints and human factors. In this study, we provide a novel approach that uses the most advanced deep learning model for object detection-YOLOv8 multi-class classifier-to identify and categorize hazards on construction sites. To achieve this, we gathered and created a curated list of construction sites photos with many kinds of dangers, like objects falling from height, Live Electrical Lines, Fire, and workers without proper protective equipment (PPE) kits. This Dataset of photos was analyzed by us and based on this we trained and assessed our model. We were able to detect multiple violations and risks within a single frame. This approach shows promising results in enhancing hazard identification and management, in construction safety practices.
Ensuring safety at construction sites is a task of great importance as it greatly impacts the health of workers and the community at large. Shocking figures from the International Labour Organization reveal that one fifth of incidents leading to the deaths of over 2 million workers annually due to work related diseases and accidents are attributed to the construction sector.
Minimizing safety risks on construction sites relies heavily on recognizing dangers but this crucial task is often hindered by various obstacles. Factors related to behavior like lack of awareness, inadequate training and limited supervision as well as issues such as fatigue, stress and distractions can significantly impact workers decision making and performance. Adding to these challenges are factors such as the unpredictable nature of construction sites along with external elements like weather conditions, limited lighting and background noise. These combined difficulties make it challenging to identify and address hazards ultimately raising the likelihood of accidents and injuries, in construction settings.
In the face of challenges there is a growing need for methods and tools to improve hazard detection and classification in the construction field.
Our study introduces a strategy to transform how hazards are identified and categorized on construction sites. We utilize technologies, like a multi class classifier and YOLOv8 a powerful deep learning model known for its object detection capabilities.
To support our approach, we carefully labeled a dataset of construction site images showing various hazards such as falling objects, electrical risks, fires and workers without proper safety gear. Using this dataset, we conducted training and evaluation of our model rigorously comparing it with existing methods. Our research focuses on assessing how well our approach performs in areas like accuracy, recall rates and speed when detecting and classifying hazards found at construction sites. The significance of our paper goes beyond academic discussions but has a profound impact on construction safety and hazard detection. Our innovative method can accurately spot hazards in an image marking a significant shift in how construction safety is managed. By providing feedback to site managers and workers our approach shows potential in improving the effectiveness of safety measures thus reducing the likelihood of accidents and injuries at construction sites. Through an examination of our method and its implications we pave the way for a future, for everyone involved in the construction industry.
The main goal of our proposed research can be summarized as follows: • • •
Developing our dataset: The quality of the training data significantly influences the
performance of a learning model. Therefore, we utilized our custom web crawler to curate a
comprising image captured by ourselves and images sourced from existing datasets. [
Object detection model: Rather than suggesting a new architecture for YOLO, we examine version 8, the most recent iteration of the program, and capitalize on its capabilities.
In recent years there has been a rise in research focused on using advanced technologies to improve
safety practices within the construction sector. These investigations explore approaches, such as
vision based surveillance and machine learning methods with the goal of boosting safety standards
which help avert mishaps and detect injuries. M. Z. Shanti et al.[
In this study a collection of data known as the CSS (Construction Site Safety) dataset has been developed to recognize individuals, vests, masks, helmets and various equipment. Stringent guidelines have been implemented to ensure the model's effectiveness in real world applications. The assessment includes the following factors:
Images with multiple class instances were preferred. The images sourced are from Google Images
where they were collected using a custom web-crawler using beautifulsoup4 Python. For picture
pre-processing, the [
We also Normalize Pixel Values of our 8-bit images by using the eq (1).
Normalized Value = !"#$#%&' )#*+' ,&'-+ (1) 255 Employing this technique stabilizes the learning process, equalizing influence by preventing larger values to dominate the learning process. Also, avoids saturation of activation functions. Another Data-Preprocessing technique we use is Data-Augmentation, used to artificially increase the size and diversity of our dataset by rotation, flipping, changing saturation and contrast of images. This helps avoid overfitting and helps the model to better generalize. We have a total of 2801 images of resolution 640x640 and 10 different detection classes (Vehicle, Machinery, Safety Vest, Safety Cone, Person, NO-safety vest, NO-Mask, NO- Hardhat, Mask, Hardhat). The Test, Validation and Train split is Test: 2605 (93%), Valid: 114 (4%), Test: 82 (3%). The data split between different classes is as follows in Table 1.
This research's main objective is to increase the accuracy of danger identification by merging preprocessing approaches for data with already-existing models. The framework of YOLOv8, which is used as the main model for PPE detection in this work, is displayed in Figure 2. The remarkable network structure of YOLOv8 converts target detection into a regression problem. In order to provide quick detection and satisfy real-time demands, it imminently offers bounding box coordinates and categorizes targets across various locations in the picture. The proposed project's methodology comprises the subsequent primary steps: • • • • •
The object detection algorithm is the Yolo-v8 algorithm.
Our datasets are utilized to train the model to identify PPE and heavy equipment. Hazards are recognized according to the equipment or machinery in their immediate vicinity and the workers' state (i.e., whether they are wearing the proper PPE).
Any infraction of safety regulations or potential hazards is identified and reported.
The YOLOv8 model is not completely trained from scratch, rather we use a technique called Transfer learning. Where the pre-trained weights (parameters) are used which were acquired on the previous dataset (COCO in our case) as a foundation. The broad knowledge of object detecting features is encoded by these pre-trained weights. We add additional layers tailored to our own dataset to replace the pre-trained YOLOv8 model's final layers (head). These additional layers are trained using annotations for the objects that we are interested in.
YOLO is an object detector that works in a single step and is implemented as a supervised learning algorithm. By using CNN once rather than repeatedly, this one-stage method is shown in Figure 3. accomplishes localization and classification concurrently—hence the name "one-stage". Cells are created on the image by forming an S × S grid. Each cell's localization and classification are handled by a CNN, which produces many potential bounding boxes and evaluates their degree of confidence. For the detection of multiple objects with their positions in a single image, the NonMaximum Suppression technique eliminates all the bounding boxes except the one which has the highest confidence for every type of object. Anchor-Free detection, a ground-breaking method that automatically predicts bounding boxes at an object's center, is incorporated into YOLOv8. Because preset anchor boxes are no longer required, the model is more durable and flexible enough to accommodate a wider range of item sizes and forms. It guarantees accurate detection results by improving object localization accuracy. An important problem with previous YOLO models was that anchor boxes frequently showed the distribution of boxes in the benchmark dataset instead of the particular distribution of boxes in a custom dataset. A variety of circumstances exist where the workers' safety is at risk. Computer vision could be of great use to us in automating the monitoring of such actions, ensuring building site safety overall. Safety conditions, which are considered as follows: • • • • •
Personal Protective Equipment (PPE): In Figure 6, PPE, which is short for personal protective equipment, is equipment used to guard against illnesses and injuries sustained at work. These can be applied to mitigate a range of occupational risks. The several PPE components are depicted in the diagram below.
Fall Protection: Utilizing protective equipment such as body harnesses and body netting to avert injuries sustained during a fall. CCTV systems can be used to keep an eye on dangerous areas and look for hazards Body Posture: Employee body posture is essential for preventing accidents, particularly in high-risk areas like scaffolding frameworks. An increased risk of events can result from improper posture and body position.
Trench Collapse: Each year, a large number of people die as a result of trench collapse. Wearing safety gear when descending into large, deep ditches can help prevent mishaps. Quality of Materials: To protect both newly constructed structures and workers, all construction and transport tools and equipment must be well-made, in excellent condition, and free from flaws.
The libraries used for this study are os, glob, pandas, numpy, matplotlib, seaborn, PIL, tqdm and Ultralytics. This experiment's primary goal is to detect things in real time and with great accuracy. Our datasets in this work were trained using the YOLO-v8 model. Table 2 shows the system needs. A single trip through the complete training dataset for learning is known as epoch. A low number of epochs leads to underfitting since there is not enough learning. Excessive epoch counts result in longer training times and the possibility of overfitting to the training set, which deteriorates test outcomes using fresh, unused data. In order to accomplish adequate training and have the best test results, the right epoch value must be chosen. But until the dataset is applied, this value cannot be determined. Consequently, the YOLO instruction suggests a 300-epoch start. Reducing the number of epochs if overfitting is observed is advised; if not, increasing it to 600, 1200, and so on when calculating the proper epoch value by assessing the learning level. The learning rate (Lr) helps determine and adjust the ratio for improving the parameter values by using the degree of loss in every learning session. Given that the loss increases with each learning cycle, this value shouldn't be set too high to avoid divergence without learning. Too low of a value causes learning to proceed very slowly. As a result, when determining an appropriate Lr, the data quality should be considered. Depending on the optimizer being used, the recommended starting Lr in YOLO varies. The default optimizer in YOLO, which is Stochastic Gradient Descent (SGD) has the value set to 0.01. However, for Adam, AdamW, and RMSProp, it is 0.001. We used the following hyper parameters Table 3. to train our model. After training our model for 100 epochs, the result we got are as follows in Figure 4. The confusion matrix for the above implementation is shown in figure 5 where the predicted and true value is mapped onto the normalized matrix elements. Our training and validation loss can be seen in figure 6.
Although the desire for progress frequently takes precedence, advancement still depends on having a workforce that is secure and well. In this study, we employed the YOLO model, version 8 (YOLOv8), as an object detection model for our dataset to detect personnel, heavy machinery, and personal protective equipment (PPE). The model shows encouraging results in correctly and real-time detecting workers, personal protective equipment, and heavy machinery on construction sites. Partially visible items in the picture and video frames could be identified by the model. Applying the created dataset to YOLO v8 suggests a consistent procedure for raising prediction accuracy through adjustments to epochs, optimizers, and hyperparameters. The testing dataset yielded the following findings for assessment metrics: precision 0.891, recall 0.797, and mAP of 0.84.4, including safety vests in the worker/PPE dataset and loaders and cranes in the heavy equipment dataset. These indicate that the model performed well in identifying underrepresented classes. The final results are competitive enough when compared to test and validation accuracies offered by other research using computer vision in construction safety management. Future work will involve improving the suggested system's ability to detect small objects. Increasing the amount of data in the training datasets will be the primary enhancement. Furthermore, hazardous situations can be foreseen and avoided with the application of spatial-temporal analysis.