MODSiam: Moving Object Detection using Siamese Networks

Abstract

Moving object detection is a challenging task in computer vision. A class agnostic model is learned to detect moving objects in a video despite their category. This is done using the proposed MODSiam that takes a single background image of the scene and the current frame as input, then the model extracts features from both inputs and merges then to output the foreground objects. A comparison of using this model with three different backbone convolutional neural networks is presented. The evaluation is done using the metrics precision, recall, F1-measure, false-positive rate, false-negative rate, specificity, accuracy, and the number of frames per second. All models are tested on the benchmark dataset CDNet, which is a dataset of videos for moving objects under different conditions like low frame rate, shadows, and dynamic background. The results show that using ResNet as a backbone produced promising results compared to other models with respect to most of evaluation metrics.