Design of a High-Speed Architecture for Stabilization of Video Captured Under Non-Uniform Lighting Conditions

Video captured in shaky conditions may lead to vibrations. A robust algorithm to immobilize the video by compensating for the vibrations from physical settings of the camera is presented in this dissertation. A very high performance hardware architecture on Field Programmable Gate Array (FPGA) technology is also developed for the implementation of the stabilization system. Stabilization of video sequences captured under non-uniform lighting conditions begins with a nonlinear enhancement process. This improves the visibility of the scene captured from physical sensing devices which have limited dynamic range. This physical limitation causes the saturated region of the image to shadow out the rest of the scene. It is therefore desirable to bring back a more uniform scene which eliminates the shadows to a certain extent. Stabilization of video requires the estimation of global motion parameters. By obtaining reliable background motion, the video can be spatially transformed to the reference sequence thereby eliminating the unintended motion of the camera. A reflectance-illuminance model for video enhancement is used in this research work to improve the visibility and quality of the scene. With fast color space conversion, the computational complexity is reduced to a minimum. The basic video stabilization model is formulated and configured for hardware implementation. Such a model involves evaluation of reliable features for tracking, motion estimation, and affine transformation to map the display coordinates of a stabilized sequence. The multiplications, divisions and exponentiations are replaced by simple arithmetic and logic operations using improved log-domain computations in the hardware modules. On Xilinx\u27s Virtex II 2V8000-5 FPGA platform, the prototype system consumes 59% logic slices, 30% flip-flops, 34% lookup tables, 35% embedded RAMs and two ZBT frame buffers. The system is capable of rendering 180.9 million pixels per second (mpps) and consumes approximately 30.6 watts of power at 1.5 volts. With a 1024×1024 frame, the throughput is equivalent to 172 frames per second (fps). Future work will optimize the performance-resource trade-off to meet the specific needs of the applications. It further extends the model for extraction and tracking of moving objects as our model inherently encapsulates the attributes of spatial distortion and motion prediction to reduce complexity. With these parameters to narrow down the processing range, it is possible to achieve a minimum of 20 fps on desktop computers with Intel Core 2 Duo or Quad Core CPUs and 2GB DDR2 memory without a dedicated hardware