Secure Deep Learning Framework for Moving Object Detection in Compressed Video

In the cloud, there is an urgent need to implement intelligent video surveillance in a privacy-preserving way. Moving object detection is an important task in the intelligent surveillance system. In this article, we propose a privacy-preserving deep learning framework to detect moving objects on compressed videos. We encrypt video bitstreams using selective video encryption to protect the private video content. We propose encrypted domain motion information (EDMI) without decryption and decompression to design three motion feature maps. Due to the sparsity of the EDMI distribution, existing convolutional backbones designed for RGB images have difficulty providing satisfactory performance. We design a novel convolutional backbone using a “subtraction” strategy to reduce model complexity. Our backbone employs residual blocks and skipping connections to reuse the EDMI at deeper layers. We evaluate our model on two large high-definition surveillance video datasets, i.e., VIRAT and Duke-MTMC. The experimental results show that the proposed framework achieves state-of-the-art detection performance compared with the most recent works. Our approach achieves an excellent privacy-utility tradeoff. Compared to previous solutions, it performs more robustly in crowded scenarios with challenges like occlusion. To our best knowledge, this is the first reported deep learning framework for moving object detection in encrypted-compressed video.