Artificial intelligence is increasingly applied in diverse domains, with unmanned ground vehicles (UGVs) serving as crucial intelligent platforms for complex task scenarios. However, the adversarial vulnerability of deep learning vision systems has become a critical constraint on operational reliability. This study systematically investigates adversarial example generation against complex background targets in UGV vision systems. We first constructed the Ground Complex Object Dataset (GCOD), containing 1,857 samples across 10 categories from ground-level perspectives, covering diverse complex environments. Three representative deep learning models—EfficientNet, DenseNet, and MobileNet—were trained on this dataset, achieving optimized recognition accuracies of 96.13%, 96.13%, and 93.92%. We implemented and evaluated two black-box attack algorithms that generate adversarial examples via only model input-output queries (resembling real-world threats). Experiments show both reduce model accuracy by over 70%, exposing severe vulnerabilities of existing target recognition models, while addressing the data gap in complex background adversarial research and revealing mainstream models’ fragility to black-box attacks.