MODY-Net: Moving Object Detection Using Multiscale Output Ensemble Y-Network

Detecting the moving objects in a video is a challenging task. Most of the currently existing methods work on a pixel level. Working on a pixel level is usually inaccurate because the pixel intensity level may change due to illumination, dynamic background, camera movement, and many other challenges. This is why, in this article, a deep network is proposed, which learns to extract features that are invariant to these challenges and then detect moving objects based on the extracted features instead of pixel intensities. The proposed network MOD-YNet consists of two main parts: an encoder and a decoder. The encoder learns to extract suitable features from both the background frame and the current frame. Then, the decoder uses the extracted features from both frames and learns to extract the objects that exist in the current frame and does not exist in the background frame. MODY-Net is evaluated on a benchmark data set called ChangeDetectionNet, using the evaluation metrics precision, recall, F1-measure, false-positive rate, and false-negative rate. The proposed model almost outperforms the state-of-the-art methods in moving object detection in all evaluation metrics by a huge gap.