基于交叉注意和跨尺度融合的车辆抛投垃圾识别  

Recognition of the behavior of throwing waste from vehicle based on cross-attention and cross-scale fusion

作　　者：陈云腾 孙振华 周杰忻 刘志[2] CHEN Yunteng;SUN Zhenhua;ZHOU Jiexin;LIU Zhi(Shaoxing Traffic Investment Group Co.,Ltd.,Shaoxing 312000,China;College of Computer Science and Technology,Zhejiang University of Technology,Hangzhou 310023,China)

机构地区：[1]绍兴市交通投资集团有限公司,浙江绍兴312000 [2]浙江工业大学计算机科学与技术学院,浙江杭州310023

出　　处：《浙江工业大学学报》2024年第6期611-620,共10页Journal of Zhejiang University of Technology

基　　金：国家自然科学基金资助项目(62073295,62072409);浙江省科技厅“尖兵”研发攻关计划项目(2022C01050);浙江省公益技术研究计划项目(LGG20F030008)。

摘　　要：旨在实时交通监控视频中智能识别违反车辆投掷垃圾(TWV)行为。TWV不仅污染环境,而且存在大量潜在危险,尤其是在高速隧道场景中,严重影响行车安全。目前,视频中TWV行为仍主要依靠人工方式检查,既耗时又费力。为此,提出了一种基于深度学习的车辆抛投垃圾识别模型(VTWIM),结合交叉注意和跨尺度融合模型(CASF)、选择性搜索和非最大化抑制(NMS),实现了基于深度剩余网络的车辆垃圾识别方法(CASF-VTWI)。首先,通过选择性搜索将一个视频帧分割为多个区域,这些区域与标有位置框的可疑对象相匹配;然后,利用CASF进行抛掷垃圾的识别训练;最后,利用NMS移除了冗余位置框,保留了最优的位置框。所提方法较好地解决了车辆垃圾的智能识别问题,对实时交通监控视频进行的实验研究证明了模型和算法的有效性与优越性。This paper aims to intelligently identify the behavior of throwing waste from vehicle(TWV)in real-time traffic surveillance video.Except for polluting the environment,there are numerous potential hazards associated with TWV,especially in the high-speed tunnel scenario,which seriously affect the driving safety.At present,the TWV behavior in the video still relies mainly on manual inspection,which is time-consuming and laborious.To address these challenges,we proposed a vehicle throwing waste identification model(VTWIM)based on deep learning.Then,combined with cross-attention and cross-scale fusion model(CASF),selective search and non-maximum suppression(NMS),we proposed a recognition method for vehicle littering using deep residual network(CASF-VTWI).Firstly,a video frame will be divided into multiple regions through selective search method.These regions will be matched the suspicious objects marked with position boxes.Then,the throwing waste from vehicle can be identified through training using CASF.Finally,the redundant position boxes are removed by NMS and the optimal position boxes are preserved.The proposed method effectively solves the problem of intelligent recognition for throwing waste from vehicle.The experimental research on real-time traffic surveillance videos has demonstrated that model and algorithm proposed are effective.

关 键 词：车辆抛投垃圾 交叉注意和跨尺度融合 交通监控视频 

分 类 号：TP39[自动化与计算机技术—计算机应用技术]