Automatic object extraction and dynamic bitrate allocation for second generation video coding

We describe a system for automatic objects extraction and dynamic bitrate allocation in a second generation object based video coder, targeting surveillance sequences. We combine model-based statistical change detection with a multiresolution-based approach for object extraction. Masks are redefined at a block level resolution and refined to improve the coding quality exploiting spatio-temporal considerations. We propose a dynamic adaptation of the quantization step of the MPEG-4 coder, based on motion information of the extracted objects. The results outperform traditional frame based coders at very low bit-rates. The masks proposed dramatically improve the coding performance, even when compared to ideally extracted masks.