IR-UWB radar based through-wall human detection with limited data via multi-hop residual CNN

Human detection from behind a wall is a modern problem that is now being studied. This has several potential applications including disaster relief, counter-terrorism efforts, and medical diagnosis. In general, radar data can be recorded, which can be used to get information about the real world. This must be turned into visuals that can be used to train a deep-learning model. We used data from a public source that was generated using IR-UWB radar to solve this problem. Using this data, we proposed a multi-hop residual CNN model for this particular case which requires minimal layers, fewer parameters, and subsequently less training time. It is achieved via multi-hop residual connections instead of usual skip connections at every alternate layer like Resnet. Even trained with 10% of total data, an accuracy of 99% is achieved within 6 epochs by our proposed model, which also converges faster when compared with Resnet 50 and Efficient Net Bl which are not able to match our achieved accuracy when trained on the same number of epochs. Moreover, lesser training time along with the reduction in trainable parameters around 2.5 times compared to VGG16 and more results of comparisons with other CNN-based architecture are also shown in this paper.

I. Introduction

Non-contact human activity detection and classification technology is important for public safety, counterterrorism, and disaster relief. For military and commercial applications, new research is developing physics-based models for ana-lyzing and imaging targets behind complex wall structures. Through-the-wall imaging uses microwave and millimeter-wave frequencies to image things or people behind walls. At search and rescue locations, through-wall radar life technology for detection is more resistant to light, noise, and climate change and is highly penetrative, precise, and anti-jamming.