Jointly Optimized Detection‐Tracking Framework With Deformable Feature Modeling and Geometry‐Aware Association for Real‐Time Worker Monitoring in Power Grid Construction Environments

ABSTRACT

Continuous worker localization and identity‐consistent tracking at power grid construction sites is challenged by equipment occlusion, extreme scale variation and visually homogeneous protective equipment. Each defeats a different stage: a worker the detector never proposes is unrecoverable downstream; overlap‐based matching collapses once occlusion drives the predicted–observed intersection towards zero; and uniform equipment suppresses the chromatic cues reidentification depends on. Detection and tracking are also tuned in isolation, so gains need not compose. This paper presents a jointly optimized detection‐tracking framework, YOLOv8‐ES + DeepSORT‐GE, for real‐time worker monitoring. The YOLOv8‐ES detector adds a C2f‐DG module combining deformable convolution with dynamic multidimensional modulation, adapting sampling geometry, kernel weighting and channel–spatial saliency within one bottleneck rather than in successive stages, dual semantic enhancement paths via transposed convolution cascades for cross‐scale reinforcement, and a dedicated P2‐level small‐target detection head. The DeepSORT‐GE tracker adds MP‐ResNet for multiscale appearance discrimination and a DeIoU geometry‐aware matching strategy with center distance penalization, its box expansion scaled by detection confidence so that uncertain predictions are gated more permissively. On purpose‐constructed power grid worker datasets YOLOv8‐ES achieves 88.1% mAP, 4.6 points above the baseline, for 1.3 M added parameters, 2.1 GFLOPs and about 3 ms per frame. The integrated framework attains 72.5% MOTA, 81.5% MOTP, 73.6% IDF1 and 66.3% HOTA, gains of 9.4, 2.4, 1.8 and 1.9 points. The upgrades prove more than additive on identity: separately 0.5 and 0.3 IDF1 points, together 1.8, an interaction supplying 56 percent of the gain. A controlled 220 kV GIS substation simulation, stratifying performance against ground‐truth occlusion duration as the captured sequences cannot, yields 73.6% fewer identity switches and 55.2% lower reidentification latency; both are conditional on the simulated appearance model and complement rather than replace the measurements on real footage. Throughput of 28.6 fps exceeds the 25 fps acquisition rate.