Thermal Object Detection Using Yolov3 and Spatial Pyramid Pooling

Abstract

Thermal Imaging refers to capturing the infrared radiation of the images and creating images out of the infrared radiation. This infrared radiation can be captured by special purpose infrared cameras. These cameras do not have the need to capture the images in ambient light. The images can be captured in darkness and can be used to detect objects and living beings. Object detection techniques earlier used for normal images using deep learning can be extended to thermal images. The paper discusses the various Convolution Neural Networks(CNN) models available to detect the objects. We then look at the YOLOv3 and Spatial Pyramid Pooling(SPP) approach to detect objects in thermal images. YOLOv3 approach to detection is revolutionary as it uses a single CNN to get the bounding boxes and class probabilities for the image. The approach discussed here contains YOLOv3 and SPP layer. YOLOv3 contains a deep convolution neural network, the convolution neural network, we then add a Spatial Pyramid Pooling(SPP) layer on top of the CNN. The SPP layer removes the need for cropping the image in the fully connected layer. This gives a speedup of over 100 times. The precision and accuracy goals of over 80 percent are met with the above approach for the FLIR dataset.