Human detection with YOLO for last-mile delivery applications using UAVs

Abstract. Low-cost alternative solutions have advanced last-mile delivery, with Unmanned Aerial Vehicles (UAVs) emerging as a promising option for logistics tasks. However, as UAV operations increasingly occur in densely crowded urban areas, safety concerns - especially for people nearby - have intensified. To ensure safe deliveries, real-time UAV path planning is essential for avoiding no-fly zones defined around individuals detected along the route. This study addresses this challenge by evaluating human detection confidence in UAV imagery using the YOLOv7 model and a custom dataset. It also estimates individuals’ positions through the Monoplotting technique. The customized YOLOv7 model achieved an average precision of 53.8% and an inference time of 9.9 ms, supporting real-time deployment. However, challenges in UAV-based human detection significantly influenced detection confidence. In one scenario, confidence exceeded 85% for individuals identified at a flight height of 30 m, while in another, the highest confidence reached 60% for imagery captured at 20 m. Despite numerous false negatives, the individual closest to the UAV was consistently detected, underscoring the applicability of the method for real-time path replanning. Regarding coordinate estimation of detected individuals, the Inertial Measurement Unit (IMU) system exerted the greatest influence on the accuracy of 3D positions obtained through Monoplotting. The phototriangulation process, which provided the sensor orientation parameters, directly impacted the results. To reduce discrepancies between estimated and actual coordinates, mathematical correction strategies may be applied; however, these require further investigation. The integration of Simultaneous Localization and Mapping (SLAM) techniques offers a promising direction for refining UAV rotation angles.