Smoky Vehicle Detection Based on Improved Vision Transformer

Abstract

The harmful exhaust emissions of fuel vehicles in the world are damaging to human health and the environment, thus detecting smoky vehicles from real road environment is significant. At present, methods of smoky vehicle detection based on deep learning have the problem of high false-positive rate. To improve the performance, a two-stage video smoky vehicle detection algorithm based on the smoke classification in the core region from detected vehicle object boxes is proposed in this paper. Specifically, the vehicle object detection is realized by the algorithm based on YOLOv3. The smoke classification is realized by combining Vision Transformer and distillation, and the loss function is optimized in the training process. Experimental results on our smoky vehicle dataset have shown that the improved model achieves an F1 score over 0.4, precision over 0.4, recall nearly 0.1 improvement compared with the basic model, which can effectively reduce the false-positive rate during detection.