FPGA-based Video Motion Tracking System

Video motion tracking is an important topic of machine vision, image processing and pattern recognition. Optical flow is an important kind of video motion tracking algorithm, which is high accuracy, and can get motion parameters of the moving target directly. But the algorithm is complex and poor real-time, so designing the real-time optical flow algorithms and corresponding hardware processing platforms is still the focus of current research.Lucas-Kanade optical flow algorithm is an important method of optical flow calculation, comparing to previous methods, which is high accuracy, simple structure and prefer to application in real-time processing. Optimizing the details of the algorithm through the Matlab is in order to trade off between the real-time and the accuracy of the algorithm. In the stage of image preprocessing, the image quantization noise is eliminated and the correlation between adjacent pixels is improved effectively by using the low-pass filter base on templates of 3D Gaussian smoothing. The optimized 3D non-Gaussian matched filter is used in the calculation of derivatives. Compared with the derivative of the traditional 2D Gaussian filter and the 3D Gaussian derivative filter, it is higher accuracy to achieve the ideal results.With the FPGA technology progressing, the speed, the internal multiplier and the internal RAM of the FPGA are increasing. Its internal resource can be allocated flexibility, and there is no limit on the pipeline stages, so it is more suitable for real-time video processing comparing with the previous DSP and PC. By this reason, the DE2-70 development system is selected as the real-time video processing platform, which has a core of the Cyclone II series FPGA, in which the calculation of LK-algorithm-based real-time optical flow is implemented. In the processing of the smoothing and derivative filters, the corresponding memory management unit (MMU) is designed, adopting the filter order of time first then space. So the capacity of the image cache and the occupancy rate of the logic cells are reduced effectively. In the floating-point computing unit, the IP core is used to implement a real-time processing. By the reasonable overall arrangement for the pipeline and the sub-pipeline, the system achieves the real-time video motion tracking for the 640*480 resolution 30 frames/s image.