In this paper, we propose a new semi-fragile watermarking scheme in the frequency domain for surveillance videos authentication. Our system starts operating by generating a binary watermark based on a novel watermark construction process. This latter combines Speeded Up Robust Features (SURF) and Maximally Stable Extremal Regions (MSER) detectors to extract frames relevant features that can resist common attacks while being fragile to intentional manipulations. Furthermore, the watermark security is improved using torus automorphism mapping. For the embedding process, Regions of Interest (ROI) are detected and then used as watermark holders. These regions are decomposed into different frequency sub-bands using Singular Value Decomposition (SVD) as well as Discrete Wavelet Transform (DWT). Then, the watermark is embedded in selected bands following an additive method. A blind detection is conducted to extract the hidden signature from the watermarked video. Evaluation results show that the proposed scheme is suitable for authentication purpose since it efficiently discriminates malicious manipulations from non-malicious ones. Besides, it preserves a high level of perceptual quality.