Architecture for Fast Object Detection Supporting CPU-GPU Hybrid and Distributed Computing

This project describes architecture for fast object detection that integrates uniform local binary patterns with convolutional neural networks (CNN) and supports CPU-GPU hybrid and distributed computing based on the Hadoop distributed computing platform. In this project we aim to extract region of interest using gradients and filters. We also use Uniform Binary Patterns from the region of interest using CPU. It also predicts accuracy using Machine learning techniques about the Uniform Binary Patterns and gets the best accuracy. We plan to use CNN to extract features using caffe Net model so that we meant to get to train the own model on our own data. Finally, we except output with bounding boxes regression on the desired locations. It is so effective for object detection to apply the CNN after fast filtering using the ULBP about Region of Interests first. The proposed architecture for fast object detection supporting CPU-GPU hybrid computing includes four phases. First, it extracts Region of Interests from an image using the binarized normed gradients (BING) which provides fast objectness estimation. Next, it rapidly and approximately checks the possibility of objectness by calculating scores using the linear support vector machine (SVM) about ULBP features extracted from suggested regions. In the third step, it verifies objectness by computing scores after obtaining CNN features from the fully connected layer (FC7) of fine-tuned Caffe Net model about confirmed regions from the ULBP. Finally, it then generates an output image with bounding boxes as a result of object detection through bounding box regression about regions with high scores. Consequently, the proposed architecture achieves 6.12 times faster processing than single node about 10K input images when using 8 data nodes.