Traffic infrared target detection in inclement weather still needs further improvement. To improve the accuracy of thermal target detection algorithms in inclement weather, this paper introduced the YOLOv4 network as detection model. By optimizing the activation function and batch size of network, it could gain lower loss, higher converge speed and good mean average precision (mAP). In our experiment, FLIR dataset and pre-trained YOLOv4 weights via MSCOCO was used to train the initial model, model with modified active function, and model with both modified activate function and network batch size. And these models were used to run the detection tests on the thermal data we shot. Through experiments the adjusted YOLOv4 model obtained a higher mAP (82.71%), lower avgLoss  (0.2012%) and a higher accuracy of rainy target detection (84%) compared with other object detectors.
 

Traffic Sensors;Inclement Weather;Thermal Infrared Image;Living Target Detection;YOLOv4 Network