GA-YOLOv7: An Improved Pedestrian Detection Method for UAV Viewpoints

Aiming at a variety of complex traffic scenarios, such as the sudden appearance or disappearance of pedestrians in the UAV view and crowd density, as well as the problems of misdetection as well as omission of mainstream detection methods, we propose an optimized YOLOv7 target detection algorithm. Comparing the performance of common target detection algorithms, we choose YOLOv7 algorithm as the basic framework, and improve the ability of the algorithm to perceive the spatial information of the feature map by replacing the CoordConv module; introduce the Gather-Excite attention mechanism to strengthen the utilization of contextual information; increase the small-target detection layer to adapt to the object targets at different scales, and improve the accuracy of the algorithm; compare the performance of different loss functions, and choose the optimal target detection method. Performance of the loss function, select the loss function Wise-IoU v2 with optimal performance, and introduce the dynamic non-monotonic focusing mechanism to accelerate the convergence speed of the algorithm. Combining the self-constructed dataset and Visdrone2019 public dataset, which is divided into training, validation and test sets in the ratio of 8:1:1, the improved GA-YOLOv7 model improves the detection accuracy by 3.1% to 91.7%, which is in line with the real-time detection requirements.