Target Detection Method for Hazardous Material Transport Vehicles in Service Areas Based on Knowledge Distillation

Currently, the safety management of hazardous materials transport vehicles in highway service areas primarily relies on manual monitoring, which is subject to limitations and susceptible to the oversight of monitoring blind spots and critical risk points. To tackle these challenges, this paper introduces a target detection method rooted in knowledge distillation. Initially, the study delves into the research subjects and challenges, establishing a large-scale dataset. Subsequently, by integrating advanced Next-ViT and AIFI modules to supplant the backbone network of YOLOv5s, the capability of the model to delineate features of hazardous materials transport vehicles is augmented. Furthermore, employing an efficient decoupled head as the detection head enhances the model's inference velocity and categorization prowess, henceforth utilizing the refined YOLOv5s as the teacher model. Ultimately, the knowledge distillation approach facilitates the improved model in guiding the YOLOv5s learning process. The experimental outcomes illustrate that the student model excels in the test set, achieving an mAP@0.5 of 94.10%, mAP@0.5:0.95 of 78.94%, and a detection rate of 79 frames per second. In comparison to the prevailing mainstream target detection models, this model exhibits superior detection precision and swifter velocity.