基于双教师联合蒸馏的黑烟识别算法

Black smoke recognition algorithm based on dual teacher joint distillation

摘要:

目的：基于监控视频的智能黑烟车识别方法可以有效节省人力和物力资源，具有广泛的应用前景。但车辆排放的黑烟具有半透明性，与背景中的沥青路面不易区分，且随着车辆运动，黑烟会产生烟羽扩散，具有不稳定的形状，最终导致黑烟识别的精确率和召回率不高。方法：首先，利用YOLOv5s-MobileNetv3模型对车辆进行定位和排烟区域截取，以降低后续处理的数据量；其次，利用K-Means算法对车辆尾部黑烟进行聚类得到具有普适性的宽高比，根据宽高比提取得到车辆的尾部排烟区；最后，提出一种双教师联合蒸馏的黑烟识别方法进行车辆尾部黑烟识别。结果：在某高速路段包括黑烟车的62段监控视频上进行训练及测试，目标车辆检测速度为76fps，在精确度94.7%的前提下，召回率为97.5%，黑烟特征识别精确率93.43%，误报率为6.52%。结论：利用轻量级网络对车辆进行定位降低了算法复杂度，保证了方法的实时性；本文提出的双教师联合蒸馏网络模型在保证较高精准率的前提下，识别时间具有明显优势。

Abstract:

Purpose: The intelligent black smoke vehicle recognition method based on surveillance video can effectively save manpower and material resources, and has broad application prospects. However, the black smoke emitted by vehicles is semi transparent and difficult to distinguish from the asphalt pavement in the background. As the vehicle moves, the black smoke will generate smoke plumes and have unstable shapes, ultimately resulting in low accuracy and recall of black smoke recognition. Method: Firstly, use the YOLOv5s-MobileNetv3 model to locate the vehicle and capture the smoke exhaust area to reduce the amount of data for subsequent processing; Secondly, the K-Means algorithm is used to cluster the black smoke at the rear of the vehicle to obtain a universal aspect ratio, and the exhaust area at the rear of the vehicle is extracted based on the aspect ratio; Finally, a dual teacher combined distillation black smoke recognition method is proposed for vehicle exhaust black smoke recognition. Result: Training and testing were conducted on 62 surveillance videos of a certain highway section, including black smoke vehicles. The target vehicle detection speed was 76 fps, with an accuracy of 94.7%. The recall rate was 97.5%, the accuracy rate of black smoke feature recognition was 93.43%, and the false alarm rate was 6.52%. Conclusion: The use of lightweight networks for vehicle localization reduces algorithm complexity and ensures real-time performance of the method; The dual teacher joint distillation network model proposed in this article has significant advantages in recognition time while ensuring high accuracy.