Lightweight UAV-based object detection and tracking for intelligent oil and gas field safety monitoring

Abstract

Continuous, reliable monitoring of geographically dispersed oil and gas production facilities is essential for accident prevention, asset integrity, and regulatory compliance, yet ground-based inspection is labour-intensive and temporally sparse. Unmanned aerial vehicles (UAVs) offer on-demand aerial coverage, but automated interpretation of production-site imagery is hindered by extreme target-scale variation, irregular equipment geometries, and the limited compute available on airborne platforms. This paper presents an integrated lightweight detection-and-tracking framework tailored to oil and gas field surveillance. The detector augments YOLOv8s with deformable C2f (DCNv2) modules for geometry-adaptive feature extraction, a progressive feature pyramid (AFPN) that suppresses cross-scale semantic conflict, a parameter-shared detail-enhanced head (LSDECD-Head), a Focaler-GIoU regression loss, and LAMP structured pruning for embedded deployment. For dynamic situational awareness, an improved ByteTrack introduces a spatial–appearance similarity matrix encoding operational-state cues and an acceleration-aware Kalman correction. On VisDrone2019, UAVDT, and a self-constructed oil-and-gas dataset (OGF-UAV), the method attains 44.1%, 57.6%, and 56.3% mAP@0.5 with only 7.8 M parameters at 78 FPS, and reduces tracking identity switches by up to 43.2%. The pruned model runs at 18.3 FPS on an NVIDIA Jetson Orin NX 16 GB under TensorRT FP16, confirming practicality for real-time field deployment.