Object Detection in Video Based on Transfer Learning Using Convolution Neural Network

Abstract

Object Detection has been an active area of research and development since the past few years and due to its diverse applications it continues to be a challenging research topic. It is indeed an evident fact that convolution neural network have shown a remarkable progress for various vision related tasks such as image classification object detection etc. In this thesis, we introduce a complete framework of object detection in a real time video using the concept of transfer learning. In this thesis the model used to train the system is a Deep Neural Net. The whole concept of how deep neural net learns to recognize the pattern in the data is termed as Deep Learning. It originates from Machine Learning which is basically a subbranch of Artificial Intelligence (AI). Deep Learning is mainly based on Artificial Neural Network with a concept to mimic the human brain i.e. Artificial Neural Networks are computational models which work similar to the functioning of human brain. Here, we use convolution neural networks to train our system on ImageNet CIFAR-10 dataset and use this trained system to detect objects in YTO dataset. We have applied various image processing techniques so as to achieve a better accuracy in comparison to the state-of-the-art methods for object detection. Basically, in this thesis, the technique that we have incorporated is transfer learning. Transfer Learning, as the name indicates is a machine learning method where the system make use of the knowledge gained while solving one problem and applying it different but related problem. Here, we are training our system with the image dataset (CIFAR-10) and we are testing it on video dataset and videos are basically collection of related and continuous images. Our framework has an accuracy of 85.99 on our own dataset and an accuracy of 61.96% on YTO dataset which is better than the state-of-art results