Embedded Multichannel Real-Time Images Processing System Based On DSP Platform

Real-time image processing has been one of the hot topics in computervision field for a long time. And recently, with the development ofautomotive electronics, systems for vehicle surrounding monitoring begin toappear. The system displays the environment around the cars to the driversusing a bird-eye view image in order to improve safety driving. Real-timeimage processing is key technology in the surrounding monitoring system.However, at the present, the researches on the image stitching are mostlybased on PC platform whereas seldom on the embedded platform especiallythe DSP system.This article is mainly focus on the research of image processingalgorithm, and proposes the basic frame of the system. Choosing SURFalgorithm in images matching, RANSAC algorithm in image stitching andweighted average method in images fusion, the article proposes an improvedsystem in image processing.On the level of core algorithm, in order to improve the efficiency of theimage processing, the study looks into three aspects in the algorithm: theweight of box filter, the sub-region selection and the Haar wavelet samplingpoint range within the sub-region. In comparing the real-time quality, thestitching accuracy and the robust quality of the algorithm basing on theresult of the experiments using variable control method, the study proposes aparameter setting strategy.On the system algorithm level, based detail analysis of the results of theexperiments, the algorithm is improved step by step. To reduce theaccumulated image location errors, the study proposes the improvement inthe images stitching process and increasing the number of rectificationimages. And aiming at the special requirements of the bird-eye view imageprocessing, the stitching sequence and weighted average methods areredesigned. These improvements increase the display effects of the finalresult and at the same time assure the calculation efficiency. After improvements, the algorithm is with high performance and can successfullyput out a complete circle of images well fused.On system implementation, this article proposes a software frameworkbased on DaVinci system of TI, a DSP platform as well as a bird-eye imagesprocessing system. In this framework, the DSP side is responsible for imageprocessing while the ARM side mainly deals with the IO devices with theV4L2driver. The two sides both have access the shared memory to swapimage data with high efficiency and the framework makes the blockingcaused by mutually exclusive lock as quick as possible. At the end, the studygives out the DSP implementation roadmap and some special considerationswith reasonable tests results.Compared to the previous researches, the study implements a bird-eyepanoramic monitoring system based on embedded DSP platform andimproves the algorithm dealing with the calculation efficiency. The studyalso proposes a parameters setting strategy of SURF algorithm which can beadapted to different images processing projects, for example, in the videostream stitching, proper parameter settings can improve30%of thecalculation efficiency while assuring the stitching accurancy. Parametersetting method permits a wider use of the SURF algorithm.