A Novel Approach on Unsupervised Dynamic Background Extraction Using Autoencoders

Abstract

In this Study a novel method has been used for extracting the background from given images. Different from the existing approaches, a Convolutional Neural Network (CNN) Autoencoder (AE) has been trained with frames produced from the same stationary camera source, paired with random frames from the same pool for each sample as a label. A little over 4000 RGB images with the dimensions of 640x480 has been used for training and around 450 of them was used for testing. The mentioned model has 4 convolutional layers each in encoder and decoder sections. The training was conducted for 500 epochs and the value of epoch loss went down to 2.13x10-3 and 2.38x10-3 for training and validation respectively. After the training of the model, generated background samples were subtracted from the input images and was turned into a binary image using two different segmentation methods: HSV Thresholding and OTSU. To use as the ground truth, test images were hand labeled. Mentioned approach had an F1-score of 62.36% for HSV Thresholding and 69.63% for OTSU methods.