Faster Region-Convolutional Neural Network with ResNet50 for Video Stream Object Detection

Detecting objects through video frames is a difficult process due to several issues, such as poor quality of frames, pose changes, occlusions, video defocusing, and motion blur. The closely arranged frames make it more complicated to detect the object in video streams. A common strategy employed to identify the problem is to correctly detect objects through video sequences. Considering these problems, the Faster Region-Convolutional Neural Network (FRCNN) is implemented to detect objects in video streams. At first, the Region proposal network (RPN) is used to generate the objects feature, and then the Convolutional Neural Network (CNN) classifies the RPN objects' features. Then the proposed architecture is combined with the ResNet50, which is the backbone of the Faster R-CNN pipeline, defining the FRCNN-ResNet50 object detector that constant detects objects in video streams. Then the model is evaluated in the DAVIS dataset and gets the F-measure 0.893, S-Measure 0.859, and Mean Absolute Error (MAE) 0.016 respectively. Compared to recent works like CSAttention-ConvLSTM, Deep-GCN, CFCN-MA, the proposed model achieves a higher detection rate and lower false detection.