A method for identifying illegal parking from unmanned aerial vehicles based on adaptive anchor boxes and multi-objective perception

With the continuous increase in vehicle numbers, traffic violations have become prevalent. Traditional detection approaches are often labor-intensive and constrained by geographical and temporal limitations. This paper introduces an innovative method for analyzing unmanned aerial vehicle imagery of traffic violations, utilizing adaptive framing and multi-objective optimization techniques. We propose an adaptive anchor box algorithm based on SAMOPSO-Kmeans, which optimizes the number and dimensions of anchor boxes suitable for vehicle datasets. This optimization modifies weight coefficients and learning factors in a multi-objective particle swarm optimization algorithm, allowing for dynamic adjustments in response to iteration counts. Expanding on the SAMOPSO-Kmeans framework, we have developed a multi-objective detection algorithm, named SAMOPSO-EfficientDet-YOLOv7, for detecting traffic violations. This algorithm integrates the YOLOv7 detection network with the EfficientDet architecture, which employs a composite scaling strategy to adjust the network's depth, width, and resolution simultaneously. The proposed framework significantly enhances performance compared to conventional methods, reducing the parameter count from 14.5 M to 9.1 M and improving mean average precision and accuracy by 1.48% and 10.52%, respectively.