Semi-Supervised Object Detection (SSOD) enhances model performance by synergistically employing a limited quantity of labeled data in conjunction with a substantial volume of unlabeled data. Nonetheless, when applying SSOD techniques to military scenarios, additional challenges arise: (1) Due to limited battlefield labels and complex environments, pseudo-label noise is amplified, and traditional filtering methods struggle to distinguish reliable predictions from those with localization bias. (2) The long-tailed distribution and dense arrangement of military objects create gradient bias, while current methods lack the adaptability to handle low-frequency categories and overlapping objects. To this end, we propose a collaborative optimization framework named MSCO-Det. First, Adaptive Gaussian Pseudo-label Assignment (AGPA) dynamically generates Gaussian thresholds through a joint classification-localization confidence model to filter noisy labels. It employs task-decoupled supervision to preserve reliable samples for classification, objectness, and regression tasks, thereby minimizing information loss. Second, Low-Risk Sample Selection (LRSS) is implemented by fusing classification probability, bounding box spatial sensitivity, and IoU reliability to construct dynamic risk evaluation indexes, in combination with adaptive K-value regulation to enhance the robustness of dense object detection. Third, Self-Regulated Class Balancing (SRCB) is employed, which leverages real-time category statistics from hybrid sample generation and gradient reweighting to mitigate the supervised attenuation issue associated with low-frequency military objects. Comprehensive experimental results indicate that the proposed method significantly outperforms mainstream semi-supervised approaches on the COBA and MAD military object datasets.