Object Detection on Embedded System

Abstract

Human detection in thermal infrared images postures a difficult challenge because infrared images have a lot of noise and they have very low resolution. We use Local Adaptive Steering Kernel (LASK) method to extract features from infrared images on the basis of image geometry concisely. We perform training less detection of pedestrians in low resolution infrared images by creating a rigid detector template. We detect humans by cross correlation in frequency domain. The cross correlation is between features of detector template and target image features. This speeds up the detection as compared to sliding window approach. The aim of this work was to implement method that can detect humans in thermal infrared images on computer and as well as on embedded system. Our implementation is performed on benchmark thermal infrared images of dataset “OSU Thermal 01 Pedestrian database”. We implement this method on dell 6th generation, core i5, 2.4 GHz with 8GB Ram on MATLAB and on single board computer ODROID Xu4. It has “Samsung Exynos5422 Cortex™-A15 2Ghz and Cortex™-A7 Octa core” alongwith 2GB Ram of “LPDDR3”. The execution of this method is computationally cost effective than sliding window approach. There are two main tasks of this thesis, first to employ an algorithm which can extract features from images and then form a feature vector and second is to design a detector based on features extractor which can classify objects from non-objects. The main input of this thesis is the first part where we have optimized an algorithm to be used on an embedded system.