| Summary: | 碩士 === 國立臺北科技大學 === 電機工程研究所 === 103 === In recent years with the growing of global economy and all kinds of science technology. All kinds of new vehicles have been produce into the market. One of the most growing topic on vehicle is the traffic safety. Most of the high end vehicle have contain all kinds of passive and active safety feature, but they are non-cost effective to majority of vehicles on the market. So in this thesis we would like to propose a cost effective system that could be ported on majorities’ vehicle.
In this thesis it propose a new method on human detection. First calculate the gradient of each pixel in the image. Using interpolation manner to quantiles gradients in 7 fixed direction vector, Accumulate to form Histogram of oriented gradients (HOG) feature. In order to strengthen features. This thesis proposes HOG 3D concept. With extra calculation on the front and rear vector, it Strengthen the overall features. To reduce the loading of SVM classifier. Import a dimension reduction method call Fisher Criterion Principal Component Analysis (FCPCA). To select the Principal Component, to achieve to goal of dimension reduction. Last using faster Sodia-ml SVM classifier to determine whether it is a human or not.
The final goal will be implement the human detection system on NVIDIA Jetson K1 (TK1) embedded platform. It contain quad ARM Cortex-A15 core with 192 CUDA GPU core. Using the technology of CUDA parallel programming, and tune the parallel code to fit TK1 embedded system beastly. To improve the accuracy and performance speed of human detection system on TK1. The method that this thesis propose have get a 93.1 accuracy. On personal computer with a NVIDIA GTX 750 ti GPU card it could process a single frame in 58ms. It prove that the method this thesis propose have both accuracy and process speed.
|
|---|
