Parallel and Accelerated Feature Extraction of Manipulative Scene of Space Dim Target

Abstract

Aiming at the problems of difficult feature extraction and high real-time requirement for weak targets in space manipulation scenes, this paper designed a parallel stream HOG feature computing architecture based on multi-cache interaction, and adopted cell stream splitting and multi-angle interval parallel computing methods to improve the degree of parallelism. The whole architecture adopts 16-bit fixed-point base calculation, transcendental function combined with polynomial expansion and table lookup method to improve the calculation accuracy. 16 detection windows at the same time, the use of parallel computing, effectively improve the response to more than $\mathbf{1024\times 1024}$ pixels high-definition camera real-time computing applications. After testing the data with low signal-to-noise ratio, the error from theoretical calculation is less than 3%. The experimental verification shows that compared with previous designs, the proposed method occupies the same number of resources except Block RAM, and the computational efficiency is increased by at least 57.2%.