Moving Object Detection for Thermal Video Using Encoder-Decoder Type Deep Learning Framework

Background subtraction (BGS) is an important stage in many thermal surveillance systems. However, the foremost objective of any such system is to detect local changes, and the system could be utilized to face many real-life situations. Further, most existing methods need to perform better in thermal video sequences. This paper presents a novel approach for detecting objects in motion for challenging thermal video scenarios. The developed technique leverages an encoder-decoder model, where the encoder integrates an altered ResNet-101 model and a feature extraction framework (FPF). The encoder employs a transfer learning mechanism on a pre-trained modified ResNet-101 network, while the FPF ensures the preservation of multi-scale and multi-dimensional features across various scales. The decoder framework is comprised of piled transposed convolutional layers responsible for translating features back to the image. To attain the effectiveness of the designed scheme, investigations were conducted, comparing it against twenty-one existing methods. The results acquired by the developed algorithm are corroborated using subjective as well as objective analysis. It may be perceived that the proposed model achieves better performance against nineteen existing techniques.