A High Efficient Multi-Scale Multi-Target Object Detection Processor Using SGI Features

Abstract

This paper describes a high efficient object detection processor we developed. Statistics of gradient images (SGI) adopted as image features can be quickly calculated from integral images. The SGI features are scale invariant and free from image resizing in multi-scale window search. We developed an object detection processor using the SGI features and a linear Support Vector Machine (SVM) classifier. This processor includes a two dimensional array of processing elements (PEs). The PE calculates a product sum of the SGI features and the SVM coefficients about a cell, a small square image as a unit of feature calculation. The PE array produces the sum of all cells in a detection window, given to the SVM classification. A dynamic reconfiguration of the PE array enables to change size and aspect ratio of the window. Various kinds of objects can be detected by rewriting the coefficient table in the PE. The PE array of a 6×6 configuration gives high detection throughput having 1 window per 2 clock cycles for face and vehicle, 4 clock cycles for pedestrian. We implemented the processor on an FPGA. The FPGA processor achieved the pedestrian detection of Full HD 30 fps resolution with multi-scale of 32 levels at 25.8 MHz operating frequency. Compared with conventional processors, the proposed processor is featured by high throughput with small circuit scale and low operating frequency.

Author

Masayuki MIYAMA
Faculty of Electrical and Computer Engineering, Institute of Science and Engineering, Kanazawa University

Read article