Vehicle Target Detection in Bad Weather Based on Infrared Thermal Image and YOLOv5

To enhance the accuracy and real-time performance of vehicle target detection in adverse weather conditions, we propose a vehicle target detection method utilizing transfer learning with YOLOv5 for infrared thermal images. First, we employ the publicly available FLIR ADAS dataset to pre-train the model. Subsequently, the weights and parameters from the pre-trained model are transferred to the private dataset, Private_data, for fine-tuning and model retraining through transfer learning. Next, we implement the TensorRT framework to optimize and accelerate the model. Experimental results indicate that this method improves vehicle detection accuracy by 0.6% compared to the original YOLOv5 model while reducing the model training time by 30%. After deploying the model with TensorRT acceleration, its size is reduced to one-third of the original, and the inference time is 6 ms, which effectively meets the real-time requirements for automatic driving target detection tasks, while the average accuracy of the model remains largely unchanged.