OBJECT DETECTION IN VIDEO USING REGION BASED CONVOLUTION NEURAL NETWORK

Abstract

Object detection is the task of recognizing and localizing objects in an image. Object detection in images have many applications including object counting, Visual Search Engine, security, surveillance etc. In this research work region based approach for object detection Faster R-CNN was refined to detect objects in videos more efficiently. The contextual information was incorporated by adding recurrent neural layer (LSTM) between consecutive frames in videos. The system was developed and evaluated using ImageNet VID Dataset for videos. The system provided mean average precision of 0.64 for detection of objects in that dataset which was higher than mean average precision of 0.56 for Faster R-CNN without LSTM. The system provided more than 15 % enhancement in F-measure of detection of objects in videos than Faster R-CNN method without LSTM layer. Also performance Faster R-CNN method based on VGGNet architecture was compared with Residual Network (ResNet). The architecture of ResNET was modified to incorporate region proposal network to detect and classify objects and their boundary in an image. The results were compared with Faster R-CNN based on VGG-16 on PASCAL VOC dataset. It was found that Faster R-CNN based on ResNET provides mean average precision of 0.78 which is better performance on PASCAL VOC dataset than VGG-16 Net architecture with mean average precision of 0.699. It was also found that ResNet based system was faster than VGGNet based system for object detection using Faster R-CNN method.