Improved YOLOv7-Tiny for Hazardous Vehicle Detection

In addressing the challenge of distinguishing between regular trucks and hazardous transport vehicles within the park, this paper introduces an enhanced algorithm for hazardous material vehicle detection, leveraging the YOLOv7-tiny model. The algorithm's objective is to swiftly recognize hazardous material markings on vehicles and differentiate hazardous material transport vehicles. Initially, a dataset comprising 8,000 images and 10,000 hazardous material vehicle samples is compiled. Leveraging prior knowledge of hazardous material markings enhances the detection accuracy of hazardous vehicles. Subsequently, the YOLOv7-tiny network undergoes several enhancements: the CoT-ELAN network structure is proposed to improve feature extraction for small targets by integrating attention mechanisms, and a decoupled head structure is introduced to address different aspects of classification and localization, thereby boosting overall performance. Experimental findings showcase the superior performance of the enhanced algorithm on the internally constructed dataset, achieving a 5.6% increase in average precision compared to the traditional YOLOv7-tiny model. With a detection speed of 70.5 FPS on GPU devices, the algorithm effectively meets the hazardous material vehicle detection needs within the park.