<p>Thermal imaging is increasingly applied for reliable intrusion detection in intelligent surveillance systems under dark environment. Yet, precise intruder posture classification (e.g., crawling, climbing) in thermal images is still a challenging task due to the low resolution, background noise, and weak spatial-spectral correlations. To the best of our knowledge, most existing methods neglect to consider the scale variations, and the inter-label dependencies, which decreases the capacity when applied to real-world scenarios. We introduce a Dual Attention Network (DAN) for multi-label thermal image classification, tailored for intrusion detection. The proposed methodology consists of a multi-scale feature extraction with ResNet-50 as the backbone to adapt with posture size variation, a Spatial Attention Module (SAM) to emphasize on posture related area, and a Channel Attention Module (CAM) to attend to inter-channel relations. The fusion layer then integrates enhance spatial and channel features to improve final predication. The proposed model was trained and evaluated on a synthetic thermal dataset (PDIWS), consisting of 2,000 training and 500 testing images representing five distant poses: creeping, crawling, stooping, climbing, and other. The experimental results demonstrate that the proposed model achieved the best results with 97.98% classification accuracy, surpassing the baselines CNNs (ResNet-50: 97.02%) and other backbone networks. Ablation experiments verified the integration of spatial attention (+ 0.33% accuracy) and channel attention (+ 0.39% accuracy). In order to improve the effectiveness, we used Grad-CAM to visualize the attention maps to confirm the model’s ability.</p>