We propose a construction site management and control technique based on drone visual surveillance. Our approach employs a cyclic drone visual image acquisition method to collect and analyze visual information from infrastructure construction sites. After extracting edge contour features from the images, we apply a complex envelope detection method to segment the site into block regions. A block-based monitoring strategy is then used to extract abnormal features, enabling the identification of unsafe and uncertain visual components. Combined with artificial intelligence analysis, our method achieves real‑time site management and anomaly detection. Simulation results demonstrate high accuracy, efficiency, and real‑time performance in recognizing abnormal states on construction sites, offering significant practical value.
In modern engineering construction, efficient management and control of infrastructure sites are essential for ensuring safety, reducing costs, and improving quality. Traditional manual supervision and video monitoring suffer from low intelligence and high subjectivity. To address these challenges, we integrate drone visual analysis technology with image processing and information processing techniques for abnormal state monitoring. By leveraging artificial intelligence classification and pattern recognition, we achieve optimized visual analysis for construction site management. A key aspect of our work is the incorporation of drone regulation principles, which govern the safe and compliant operation of unmanned aerial vehicles (UAVs) in construction environments. Adhering to drone regulation ensures that our visual surveillance system operates within legal and safety frameworks, minimizing risks and maximizing operational efficiency.
Visual Information Acquisition and Preprocessing
Drone Visual Image Acquisition
We capture visual information from construction sites using a cyclic drone flight pattern. The acquisition process employs edge‑block segmentation combined with horizontal and vertical scanning techniques. The block segmentation model is illustrated in the following figure (the actual image is referenced via the provided link).
The acquisition procedure consists of several steps. First, we center on edge pixel points to collect missing information regions in the drone visual image. For a center pixel point p, we define a block region Ψp around it. Missing texture regions and abnormal state areas are linearly superimposed to construct the pixel value distribution field model:
$$
S = \{ s = (x,y) \mid 1 \le i \le M,\; 1 \le j \le N \}
$$
By combining vertical and oblique scanning, we obtain the drone visual image. The site region confidence and template matching terms are given by:
$$
C(p) = \frac{\sum_{x \in (\Phi \cap \Psi_p)} I(x)}{|\Psi_p|}
$$
$$
D(p) = \frac{\nabla I_p^\perp \cdot n_p \cdot \cos \alpha}{M}
$$
Here, x denotes the pixel point in the intersection of the acquisition module Ψp and the information expression region Φ. Based on these acquisition results, we proceed with edge contour feature extraction and region segmentation for site management and feature recognition.
Edge Contour Feature Extraction
After image acquisition, we perform edge contour feature extraction using binary transformation. The edge pixel estimate for drone visual imaging is:
$$
\hat{\eta} = \arg\min_{\eta} L(\eta), \quad \hat{R}_x = \frac{1}{K} \sum_{k=1}^{K} x_k x_k^H
$$
The affine invariant moment for edge feature segmentation of the construction site management image is:
$$
m_{pq} = \sum_{m=1}^{M} \sum_{n=1}^{N} x^p y^q f(x,y)
$$
where xp and yq are coordinates in the invariant moment feature space. Using scale‑invariant moment estimation, we obtain the uniform pixel‑level intensity:
$$
f_R(z) = \begin{pmatrix} f_x(z) \\ f_y(z) \end{pmatrix} = \begin{pmatrix} h_x * f(z) \\ h_y * f(z) \end{pmatrix}
$$
Here f(z) is the pixel intensity of the input drone visual image, and * denotes convolution. The grayscale histogram of the drone visual image is:
$$
P(X=x \mid Y=y) = Z^{-1} \exp\left(-U(x,y)\right)
$$
where U is the pixel intensity of the edge feature distribution function, and Z is the grayscale distribution coefficient. Assuming the site management region image size is m×n, we perform a cyclic block traversal search with O(mn/a²) iterations. The priority coefficient P(p) for the detection block Ψp is:
$$
P(p) = C(p) \times D(p)
$$
P(p) consists of the correlation matching values of C(p) and D(p). We then binarize the image and apply median filtering on the edge parts to reduce noise intensity. Subsequently, we use a complex envelope detection method for block region segmentation of the construction site, improving the accuracy of abnormal state detection.
Optimized Implementation of Construction Site Visual Management
After cyclic drone visual image acquisition and edge contour extraction, we proceed with abnormal feature extraction and state recognition for real‑time site management. We employ a complex envelope detection method for block region segmentation and a block‑based monitoring method for abnormal feature extraction. The detection block size Ψp is determined based on the template size in the site region. When an abnormal feature is found in the chrominance component, we perform seed point matching. Within the complete information region Φ, we define the block segmentation area Ψp and the block density Ψp‘. To highlight texture pixel information, we use pixels with high statistical features for edge template matching. Assuming the pixel sample block sequence of the management image satisfies Ψq (Ψq ⊂ Φ), we perform block segmentation on the drone visual image using sample template covering method in four local directions (horizontal, vertical, diagonal, anti‑diagonal). The abnormal uncertainty feature extraction result is:
$$
\Psi_p’ = \arg\min_{\Psi_q \subset \Phi} d(\Psi_p, \Psi_q)
$$
where Ψq is the frame sequence of the current image block region, and d(Ψp, Ψq) represents the difference in monitoring features between normal and abnormal images. On the RGB components of the management region drone visual image, the feature point matching template function within the confidence interval is:
$$
EyeMapC = \frac{1}{3} \left( C_b^2 + \bar{C}_r^2 + C_b / C_r \right)
$$
where Cb2, Ȼr2, and Cb/Cr are ratios of prior pixel and pixel fusion. Using binary segmentation for corner detection in local block regions, we obtain the joint probability distribution for accurate detection of abnormal parts on the construction site:
$$
P(y_{w3} \mid x_{w3}, \theta, \beta) = \frac{1}{Z(\beta_i)} P(y_{w3} \mid x_{w3}, \theta) (y_{w3} \mid \beta_i)
$$
Here Z(βi) = Σyw3 P(yw3 | xw3, θ) ( yw3 | βi ) is a partition function. When u ∈ N(w) \ w‘, we set yki = 1 in the confidence region; otherwise yki = 0, i.e., yi = [yi1, yi2, …, yik, …, yiK]. Based on these discriminant functions, we use block monitoring to extract abnormal features and analyze unsafe and uncertain visual components on the construction site. Finally, we input the feature extraction results into an artificial intelligence recognition system, employing a BP neural network classifier for feature classification and recognition, thereby achieving real‑time site management and anomaly detection.
Simulation Experiments and Performance Analysis
To evaluate the performance of our method for construction site monitoring and management, we conducted simulation experiments using Matlab R2011a and Visual C++ co‑programming. The parameter settings are summarized in Table 1.
| Parameter | Value |
|---|---|
| Optical flow field intensity N(w) | 3.0 |
| Block region segmentation size | 400 × 400 pixels |
| Edge template size | 42 pixels |
| Scale coefficient for block region segmentation | 0.34 |
| Neighborhood size ω for drone visual image | 7 |
The image quality of the output drone visual image is measured by the Peak Signal‑to‑Noise Ratio (PSNR):
$$
\text{PSNR} = 10 \times \log_{10} \left( \frac{\text{Peaksignal}^2}{\text{MSE}} \right)
$$
where Peaksignal is the maximum grayscale value of the central pixel in the drone visual image, and MSE is the mean square error of feature localization. Under these simulation environments and parameter settings, we performed construction site management image acquisition and monitoring. The captured visual images were processed for edge contour segmentation and feature extraction, focusing on unsafe and uncertain visual components. The output image after processing showed strong feature representation capability and high image quality, thereby improving real‑time site management and analysis capabilities.
To quantitatively evaluate the performance, we compared our method with traditional approaches. Table 2 lists the PSNR values for different methods. Our method achieves the highest PSNR, indicating stronger mining capability for site management information and improved management intensity.
| Method | PSNR (dB) |
|---|---|
| Traditional method A | 28.3 |
| Traditional method B | 31.7 |
| Our method | 36.5 |
We also compared the timeliness of different methods for construction site management. The time consumption results are presented in Table 3. Our method exhibits the shortest processing time, demonstrating superior real‑time performance.
| Method | Time (s) |
|---|---|
| Traditional method A | 2.15 |
| Traditional method B | 1.83 |
| Our method | 0.92 |
Further analysis shows that the integration of drone regulation guidelines ensures that our visual surveillance system operates within safety and compliance boundaries, reducing the risk of accidents and legal issues. The cyclic image acquisition method, combined with complex envelope detection and block‑based monitoring, effectively identifies abnormal states such as unsafe worker behaviors, equipment malfunctions, and structural irregularities. The use of a BP neural network classifier provides robust recognition of these anomalies.
Our approach not only improves the accuracy of anomaly detection but also enhances the overall efficiency of construction site management. The experimental results confirm that the proposed method outperforms traditional techniques in terms of both quality and speed. The adherence to drone regulation ensures that the system can be deployed in real‑world construction environments without violating aviation or safety rules.
Conclusion
We have developed a construction site management and control technique based on drone visual surveillance. By combining cyclic image acquisition, edge contour extraction, complex envelope detection, block segmentation, and artificial intelligence classification, our method achieves high‑accuracy anomaly detection and real‑time site management. The incorporation of drone regulation principles guarantees safe and compliant operation, making the system applicable to various construction scenarios. Simulation results demonstrate excellent performance in terms of image quality, processing speed, and detection accuracy. Future work will focus on optimizing the neural network classifier and expanding the system to handle more complex site conditions while maintaining strict adherence to evolving drone regulation standards.