The integration of deep learning techniques with thermal videos has revolutionized vehicle detection, enabling precise identification in various lighting and environmental conditions. In this study, we develop an aerial thermal dataset, which includes over 13,500 vehicle annotations using oriented bounding boxes (OBBs). This dataset is compiled from four sites across two locations in Massachusetts, USA. We also introduce an improved model for detecting and tracking vehicles in thermal videos captured by drones. We propose modifications to the YOLOv8 model, including the substitution of standard convolution functions with GSConv and an enhancement of the neck to integrate higher resolution information, specifically targeting the detection of densely packed and small OBBs in thermal images. Additionally, we introduce a selective-SAHI technique to enhance scalability and detection quality. These improvements collectively yield a 6.3% increase in the mAP50 score on our nighttime thermal dataset. Github link: https://github.com/z00bean/ThermalDroneVehicleOBB.
Analyzing Highway Work Zone Traffic Dynamics via Drone Thermal Videos and Deep Learning
2024-09-24
3561191 byte
Conference paper
Electronic Resource
English