Visual Object Recognition from Videos ( VORV )

Abstract

Object detection has many applications in computer vision, including image retrieval and video surveillance. The idea of object detection was presented with feature description and it outperformed existed algorithms and methodologies in pedestrian detection and it was a huge breakthrough in image processing field. In the past when it was required to find out a specific object from a video, one has to manually sit and watch out the complete video to take out the required part of that video manually after a lot of effort and time. As we have huge amount of data in the form of videos and to search something from videos require lots of time i.e. to go through the whole video, therefore we need to have an automated system to search specific content from videos. Videos are basically sequence of frames and frames are sequence of images. Therefore, we need to extract frames to recognize the objects in videos. Using the transfer learning technique on S SD and inception-v 2 and using Constitutional Neural Network we have trained our system to detect and recognize buildings. We have developed a system named as “Visual Object Recognition from Videos”, which recognize buildings from videos. We have fine-tuned CNN model on our dataset. As we know for Deep learning techniques, we need a huge amount of data for training to get more accurate results, therefore initially we collected more than 120 unique images of each class. Front-end is based on .NET framework C# to make it easy to use system for users. For recognition of object, we have used object detection model with Tensor Flow and also the inception-v 2 classifier, which recognize buildings in videos. For fast processing system, we have used GPU because training on images and then processing videos is very slow on CPU. The system proved precision nearly 93.26%. This is a problem solving desktop application to give ease to users of any regulatory body.