Aiming at the traditional edge detection algorithm implemented on FPGA , there are problems of lack of adaptivity and poor real-time performance , combined with the characteristics of high-speed parallel processing data on FPGA , an improved edge detection algorithm is proposed. The algorithm firstly applies bilateral filtering to the Canny edge detection algorithm , which retains the gradient information of the image edge while reducing the noise ; secondly , it uses the improved Sobel operator to calculate the gradient magnitude in four directions to improve the localization ac- curacy of the image edges ; finally , it proposes an adaptive threshold selection strategy based on the change of the first -order derivatives of the gradient histogram in order to improve the adaptive capability of the algorithm. The experi- mental results show that the peak signal-to-noise ratio ( PSNR) of the image after the bilateral filtering process is im- proved by 109% compared with the traditional Canny algorithm , and the single edge response and connectivity of the edges are significantly improved , while meeting the real-time requirements. Provide a feasible solution for the applica- tion of Canny algorithm to embedded machine vision.