Study And Implementation Of GPU Acceleration On Local Feature Algorithm SURF

Accelerating image processing algorithms by GPU is a research hot area. GPU’shigh bandwidth and many-core architecture is suitable to accelerate image processingalgorithm. Feature detection algorithms are an important part of image processing andare widely applied in Smart Surveillance, target tracking and image retrieval etc.Local invariant feature algorithms are applied in various disturbed situations becauseof its invariance of affine, illumination and rotation. SIFT, MSER and SURF arepopular local invariant features at present. Although acceleration methods are adoptedin these algorithms, their high computational complexity makes the performance oftheir PC implementation fall short of real time requirement, which confines theirapplication.This thesis discusses how to speed up SURF by GPU. The platform used ishigh-end GPU GTX480of nVidia Corporation. In this thesis, the SURF algorithmsare organized into several modules: image integration, feature detection, orientationassignment and feature description. The work is done through exposing as muchinherent parallelization as possible in original algorithms. Moreover the data layoutand data path are optimized to utilize the GPU memory resource appropriately. Thespeed of GPU SRUF proposed is about80times that of the original sequentialimplementation running on Quadro8300@2.5GHz, and its overall performance isclose to its counterparts. The proposed GPU accelerated image integration is2timesfaster than the current state-of-the-art methods. The proposed GPU SURF detector isclose to other GPU implementations. The proposed GPU SURF orientationassignment is far faster than other GPU implementations. The detection anddescription time of800X600and4000features by proposed GPU SURF is about7ms,which can fully meet real time needs.