摘要
针对激光透窗低质量图像的人体姿态识别,现有算法存在识别精度低以及严重的漏检、误检问题,鉴于此,本研究提出了一种高效的人体姿态识别算法YOLO-TCpose。设计了一种新的卷积模块,该模块可在保留全局信息的同时提升局部特征的提取能力,解决了标准卷积的固有缺陷问题;对特征融合网络进行重构,采用跨层级联和模型剪枝等方式实现浅层信息和深层信息的交互融合,以提升模型对小目标姿态的识别能力;构建了增强检测一体化网络,结合改进的ADNet图像增强去噪算法降低噪声对姿态识别的影响,提高了模型的检测精度。最后,编写人体姿态识别算法,将模型部署在Jetson NX移动开发平台上,设计了一套完整的机载激光透窗成像人体姿态识别系统。实验结果表明:YOLO-TCpose算法具有较强的鲁棒性和泛化能力,具有较高的实际应用价值。
Objective Laser throughwindow imaging technology,an advanced detection method,can effectively penetrate window glass and visualize indoor targets behind the window,providing many application prospects.In antiterrorism and stability maintenance scenarios,a throughwindow scope enables the capture of accurate information regarding the number and posture of terrorists outside the window.In traffic monitoring applications,this technology enables the assessment of a driver’s status without requiring the driver to exit the vehicle,thereby improving traffic management efficiency.However,the practical application of laser throughwindow imaging technology faces several challenges.Image quality and accurate capture of target information behind windows are significantly affected by factors such as natural illumination,object occlusion,and strong reflections from the window glass.Accurately detecting human targets and identifying their poses in complex environments is highly challenging.Conventional image processing techniques often cannot achieve accurate and efficient detection results when faced with disruptions such as changes in illumination or occlusion.Addressing these challenges requires the development of more robust object detection and attitude recognition algorithms that can be effectively implemented on edge computing platforms to meet realtime requirements.This study is highly significant,with the potential to substantially enhance fields such as antiterrorism measures,security operations,military reconnaissance activities,and traffic management.Methods Currently,laser throughwindow imaging data are not publicly accessible.Therefore,a new dataset was constructed using a laser rangegating imaging system that covers two types of scenes:natural and manmade.The natural scene includes various simulated human postures for data collection,whereas the manmade scene incorporates diverse types of glass,throughwindow distances,lighting conditions,and occlusions to enhance data diversity.Existing algorithms for human postural recognition in lowquality laser throughwindow images typically exhibit suboptimal accuracy,which is characterized by significant missed and false detection.Thus,this study used YOLOv8nPose as the base model with a targeted optimization design to address these problems.A novel convolution module was developed to improve the feature extraction ability in lowquality image scenarios with laser throughwindows,while crosslevel association and a model pruning method were used to reconstruct the feature fusion network.This approach aims to reduce the model size and improve the recognition of small target human poses.Additionally,an enhanced detection integration network that combined image denoising and postural recognition tasks enabled endtoend integrated training,further enhancing the model detection performance.Finally,a human posture recognition algorithm was implemented by deploying the model on the Jetson NX mobile development platform,creating a fully functional airborne laser throughwindow imaging human posture recognition system.Results and Discussions This study compared the performance of Faster RCNN,Alphapose,Openpose,HigherHRNet,YOLOv5s6-pose,and YOLOv8nPose algorithms for human pose recognition(Table 2).The results indicate that the YOLOv8nPose model outperforms Faster RCNN,Openpose,and HigherHRNet.Alphapose and YOLOv5s6-pose exhibit slightly better performance indicators than YOLOv8nPose.However,they significantly lag behind YOLOv8nPose in terms of inference speed and model size.Nevertheless,the proposed YOLOTCpose algorithm performs exceptionally well across various performance indicators.Additional experiments were conducted using the Openpose,Alphapose,and YOLOv8nPose algorithms in artificial and natural scenes to assess the effectiveness of the YOLOTCpose algorithm.In artificial scenes(Fig.6),comparative experiments involving single and multiple people with occlusion demonstrate that YOLOTCpose outperforms Openpose and Alphapose by achieving accurate key point positioning and significantly reducing missed detections during multiperson pose recognition.Notably,YOLO-TCpose exhibits significant advantages, particularly in scenarios involving multiperson occlusion. In natural scenes (Fig. 7), the experimental results indicate that during posture recognition tasks such as crawling during the day, standing at night, or squatting on rainy day;YOLOTCposeaccurately detects human target along with their corresponding key points, outperforming other algorithms by a significant margin. Finally, YOLOTCposeexhibits superior detection accuracy, stability, and adaptability in various environments compared to current mainstream algorithms.Conclusions This study introduces YOLOTCpose,an efficient and lightweight human posture recognition algorithm designed for detecting human poses in lowqualitylaser throughwindow images. To address the limitations of traditional convolution, a novel convolutional module was developed to improve feature extraction capabilities. Additionally, the feature fusion network was restructured by eliminating large target detection layers and incorporating small target detection layers. This adjustment facilitates the effective fusion of shallow and deep information through crosslayerconnections, thereby improving the recognition performance for small targets. By incorporating an improved ADNet denoising algorithm, an integrated network for image enhancement and pose recognition was developed, which significantly improves the detection accuracy. The experimental results demonstrate that YOLOTCposeachieves improvements of 19.3 and 26.6 percentage points in the precision and recall rate, respectively, for object detection. The mean average precision (mAP) at 0.5 and mAP at 0.5: 0.95 for keypoint detection are enhanced by 16.0 and 10.1 percentage points, respectively. In addition, the inference speed is increased by 5.1 ms, and the model size is reduced by 1.69 MB. Furthermore, algorithms for recognizing three postures—standing, squatting, and crawling—were developed, and the model was successfully deployed on the Jetson NX mobile development platform, establishing a fully functional airborne laser throughwindow imaging human posture recognition system.
作者
伍智华
程江华
刘通
蔡亚辉
潘乐昊
Wu Zhihua;Cheng Jianghua;Liu Tong;Cai Yahui;Pan Lehao(College of Electronic Science and Technology,National University of Defense Technology,Changsha 410073,Hunan,China)
出处
《中国激光》
北大核心
2025年第6期260-271,共12页
Chinese Journal of Lasers
基金
国防科技大学自主创新科学基金(24-ZZCX-JDZ-11)。
关键词
激光透窗
卷积运算
小目标检测
姿态识别
图像去噪
laser throughwindow imaging
convolution operation
small object detection
posture recognition
image denoising
作者简介
通信作者:伍智华,18627599098@163.com。
