Detection of Human Intrusion Under Challenging Situations Though Rgb-Images Using Deep Learning

Detection of an intrusion attempt by a hostile human is of utmost importance to Indian Air Force and other such installations of national importance. Intrusion detection under such circumstances often faces multiple challenges including low visibility, smoke or fog, adverse weather conditions like rain and use of camouflage by intruder. Although rich literature is available on the subject of object detection and background subtraction in general, the area of camouflaged target detection is yet to be explored in similar detail. One of the main reasons for limiting the research in this area is the lack of suitable datasets. In the aforesaid scenario, the presented work makes the following additions. Firstly, a new dataset was compiled consisting of 3500 images having humans camouflaged in different kinds of scene backgrounds. Secondly, a methodology using dilated convolution for enlargement of receptive field has been presented for segmentation of human intruder camouflaged well within terrain background by using RGB images. The proposed methodology yielded promising results in comparison to popular baseline models. Additionally, the performance of the methodology was also tested on the images under low light conditions and demonstrated encouraging results.