Abstract: Feature extraction and description are the foundation for many computer vision applications.Due to its high dimensional computation of pixel-wise processing,feature extraction and description are computationally intensive with poor real-time performance.Thus it is hard to be used in real-world applications.In this paper,the common computational modules used in feature extraction and description,pyramidal scheme and gradient computation were studied.The method used to compute these modules in parallel based on NVIDIA GPU/CUDA was introduced.Furthermore,computational efficiency was improved by optimizing memory accessing mechanism for global,texture and shared memory.Experimental results show that a 30x speed-up is obtained by GPU-based pyramidal scheme and gradient computation against that of CPU.By employing these GPU-based optimization techniques into HOG (Histogram of Gradient) implementation based on GPU,it obtains a 40x speed-up against that of CPU.The method proposed in this paper is of significance for implementing fast feature extraction and description based on GPU.

Key words: Image pyramidal scheme,Image gradient computation,GPU,CUDA