| 曾拓程,王佳俊,王晓玲,张雨诺,康栋.大场景视频监控下大坝运输车改进多目标多视觉卸料识别模型研究[J].水利学报,2023,54(5):519-529,540 |
| 大场景视频监控下大坝运输车改进多目标多视觉卸料识别模型研究 |
| Research on improved multi-target multi-vision unloading identification model of dam transport vehicle under large scene video surveillance |
| 投稿时间：2022-11-04 |
| DOI：10.13243/j.cnki.slxb.20220910 |
| 中文关键词: 卸料识别 大场景视频监控 多目标跟踪 关键点检测 细粒度分类 |
| 英文关键词: unloading identification large-scene video surveillance multiple objects tracking key point detection fine-grained classification |
| 基金项目:国家自然科学基金雅砻江联合基金项目(U1965207)；国家自然科学基金青年科学基金项目(52009089) |
|
| 摘要点击次数: 9344 |
| 全文下载次数: 629 |
| 中文摘要: |
| 运输车卸料识别对保障大坝施工安全、优化运输配置具有重要意义。然而，大坝施工中多采用全球导航卫星系统（Global Navigation Satellite System，GNSS）进行运输车活动状态分析，存在设备成本高、识别精度低等缺点。目前，通过布设高清摄像头进行坝面大场景施工监控成为趋势，但尚未见基于机器视觉的运输车卸料识别研究。针对上述问题，提出大场景视频监控下大坝施工运输车卸料识别改进多目标多视觉任务模型。首先，采用ByteTrack实现大场景监控视频中多辆运输车目标检测与追踪，记录其行驶轨迹；其次，High Resolution Net（HRNet）被用于运输车头部和尾部的关键点检测，进而结合行驶轨迹判断运输车前进、停止和后退等行进状态；再者，通过Destruction and Construction Learning(DCL)细粒度分类方法判断运输车料斗的抬升状态；最后，结合ByteTrack、HRNet和DCL的多目标多视觉任务的分析结果判定运输车卸料状态。以两河口施工视频监控为例进行验证，提出的卸料识别模型在34帧时卸料状态识别平均准确率为87.3%，卸料时间判断精度为90.3%，验证了本模型的可行性和有效性。 |
| 英文摘要: |
| Unloading identification of transport vehicle is of great significance for ensuring dam construction safety and optimizing transport configuration.However，in dam construction，the Global Navigation Satellite System(GNSS) is often used to analyze the activity state of transport vehicles，which has the disadvantages of high monitoring cost and low identification accuracy.At present，it is a trend to deploy high-definition cameras to monitor the construction of large scenes on the dam surface，but there is no research on the identification of unloading of transport vehicles based on computer vision.To solve the above problems，a ByteTrack-HRNet-DCL unloading identification model of dam construction transport vehicle under large scene video monitoring is proposed.Firstly，ByteTrack is used to realize the visual detection and tracking of multiple transport vehicles in the monitoring video，and record the driving track.Secondly，High Resolution Net(HRNet) is used to detect key points such as the head and tail of each transport vehicle，and then judge the forward，stop or backward travel status of the transport vehicle based on the track recorded.Furthermore，the lifting state of each hopper is judged by Destruction and Construction Learning(DCL) fine-grained classification method.Finally，the unloading state of the transport vehicles is determined based on the analysis results of these three visual methods.The experimental verification was carried out on the construction site of Lianghekou in Southwest China.The proposed ByteTrack-HRNet-DCL unloading identification model has an average accuracy of 87.3% for unloading judgment and 90.3% for unloading time under the average frame rate of 34，which verifies the feasibility and effectiveness of this model. |
|
查看全文
查看/发表评论 下载PDF阅读器 |
| 关闭 |