Vehicle Smoke Synthesis and Attention-Based Deep Approach for Vehicle Smoke Detection

Abstract

Third world countries are suffering from extreme vehicular air pollution due to dominating number of fossil fuel-driven vehicles on the road. Therefore, in these countries, automatic surveillance systems are in high demand for close monitoring to identify and penalize vehicles emitting excessive smoke. In recent times, deep learning-based computer vision systems are rigorously working on the same. Their accuracy strongly depends on the number of the training images taken under various imaging conditions. However, there are very few publicly available vehicle smoke datasets that could be used for training purposes. To capture on-road videos for the creation of a dataset is another challenging and time-consuming task. To aid and enhance the vehicular smoke monitoring system, in this article, we propose, a holistic dual-level framework for dataset enhancement by smoke generation along with a transformer network for efficient identification. We have created a realistic vehicle smoke generation algorithm using a range of mask patterns and filtering, which helps us to train our deep neural model by generating sufficient synthetic data. We have also proposed a transformer network on the YOLOv5 backbone, which efficiently identifies the smoke region and the smoky vehicle from the image frame simultaneously. We have shown that the lambda-implemented attention-based detection network outperforms the other state-of-the-art techniques on three baseline datasets. Sample demo videos are available at the link <uri xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">https://github.com/srimantacse/VehicleSmoke</uri> .