Design And Implementation Of Embedded Fisheye Video Correction Solution

In recent years, fisheye cameras are widely used in video surveillance and other fields with the advantage of wide angle of view, but the amount of distortion correction calculation is huge and embedded devices are difficult to achieve real-time performance due to frequency limits. While parallel computing technology can break frequency limit to improve performance.This thesis designs a multi-pixel parallel processing fisheye video correction solution with the consideration of fisheye video correction characteristics and Android OpenCL platform. This thesis extracts fisheye video correction key operator:vector scalar product and analyzes its memory access patterns. This thesis studies linear relationship between quantity of memory access and resolution. Cache capacity restrictions causing performance periodically decline is also studied. According to the relationship, this thesis assesses the OpenCL kernel performance quantitatively and concludes that four-pixel processing can achieve the best performance. For the vector scalar product, this thesis also designs code generation tool which can automatically generate the OpenCL kernel code.The test results show that the performance of fisheye video correction can achieve 35fps. Using single-pixel OpenCL processing method, the performance is 1.90 times improved compared with NEON method and 12.38 times improved compared with not using any parallel computating method. Using four-pixel OpenCL processing method, the performance is 68% improved with merely 0.8511 root mean square errors compared with single-pixel OpenCL processing method.Results of this thesis can be used for real-time fisheye video correction. For the algorithms meeting the vector scalar product characteristics, this thesis is worthy for performance evaluation and kernel design.