Efficient Object Detection and Classification of Ground Objects from Thermal Infrared Remote Sensing Image Based on Deep Learning

Wild searching and nature reserve monitoring are formidable tasks. In order to relieve the current pressure of general manpower observation, drone aerial surveillance using visible and thermal infrared (TIR) cameras is increasingly being adopted. Automatic data acquisition has become easier with advances in unmanned aerial vehicles (UAVs) and sensors like TIR cameras, which enables executives to search and detect ground objects at night. However, it’s still a challenge to accurately and quickly process the large amount of TIR data generated from this. In response to the above problems, this paper designs an enhanced ground object detection network (UAV-TIR Retinanet) for the UAV thermal imaging system. The network uses the Retinanet as infrastructure, extracts shallow features according to the characteristics of thermal infrared remote sensing images, introduces an attention mechanism and adaptive receptive field mechanism. The method achieves the best speed-accuracy trade-off on the dataset, reporting 74.47% AP at 23.48 FPS.