Secure Moving Object Detection Transformer in Compressed Video with Feature Fusion

Moving Object Detection (MOD) is essential for surveillance videos, but the large data volume necessitates significant computational resources. Cloud computing offers a solution, though it raises privacy concerns. In this paper, we propose a secure MOD framework based on the Detection Transformer. The user-side protects video data using format-compatible selective video encryption before storing it in the cloud. In the cloud, we design a method to extract encrypted domain motion information from encrypted videos. Due to the dual constraints of compression and encryption, the extracted information is highly sparse, and common deep learning methods do not perform well. To address this issue, we propose a Convolutional Neural Network (CNN)-Transformer feature enhancement-fusion method to achieve effective feature alignment and fully exploit deep-level motion information. Specifically, we modulate CNN features at multiple scales using Transformer features, where high-level semantic information and low-level spatial information are fused for accurate moving object localization. We evaluate our model on the VIRAT and DUKE-MTMC datasets, demonstrating better detection performance and greater robustness in challenging scenarios compared to previous secure MOD methods.