面向智能视频监控的空中交通管制员图像分割

Image segmentation for intelligent video surveillance of air traffic controllers

为解决复杂场景下空中交通管制员检测与分割精度低、鲁棒性差的问题,提出一种基于掩码区域卷积神经网络(Mask Region-based Convolutional Neural Networks, Mask R-CNN)的管制员图像分割模型ATC Mask R-CNN(ATC Mask Region-based Convolutional Neural Networks)。首先,构建管制员监控图像数据集(ATC Monitor Image Dataset, AMID)并用于模型训练、测试;其次,在主干网络中引入瓶颈注意力模块(Bottleneck Attention Module, BAM)以增强管制员特征提取,采取改进的柔性非极大值抑制算法(Soft Non-maximum Suppression, Soft-NMS)替代NMS算法进行候选框选取,提高对遮挡目标的检测分割;最后,基于AMID进行管制员图像分割试验。结果显示:ATC Mask R-CNN的精确率、召回率和平均精度分别为96.49%、95.62%和88.84%,表明了该方法的有效性。与Mask R-CNN相比,ATC Mask R-CNN有效降低了复杂场景的不利影响,更适用于管制员工作场景,可以为管制大厅安全管理自动化应用提供技术支撑。

Accurate detection and segmentation of air traffic controller instances from surveillance video images is the basis for behavior recognition and condition monitoring of individual controllers.To solve the problem of low detection and segmentation accuracy and poor robustness of air traffic controllers in complex scenes,a controller image segmentation model ATC Mask R CNN based on Mask Region⁃based Convolution Neural Networks(Mask R CNN)is proposed.Firstly,using the actual surveillance video of an air traffic control office for one week as the data source,the ATC Monitor Image Dataset(AMID)was constructed for model training and testing.Secondly,the Bottleneck Attention Module(BAM)is introduced in the backbone network to enhance the feature extraction of controllers.This module constructs a hierarchical attention mechanism similar to the human perception process by integrating spatial attention branches and channel attention branches.Further,low⁃level image features such as background textures are denoised.And to improve the detection and segmentation of occluded targets,the improved Soft Non⁃Maximum Suppression(Soft NMS)algorithm was adopted instead of the NMS algorithm for candidate region selection.Compared with the NMS algorithm,the improved Soft NMS algorithm adopts a new intersection⁃union ratio calculation method and sets an attenuation function for the suggestion box where the intersection⁃union ratio is greater than the threshold.Therefore,adjacent occlusion targets can be avoided from being mistakenly deleted.Finally,the controller image segmentation experiment is carried out based on AMID.By selecting relevant evaluation indicators to evaluate the performance of the model,the precision,recall,and average precision of ATC Mask R CNN are 96.49%,95.62%,and 88.84%,respectively,which proves the effectiveness of the proposed method.The experimental results show that compared with Mask R CNN,the proposed method effectively reduces the adverse effects of complex scenes,is more suitable for controller work scenarios,and can provide technical support for the application of safety management automation in the air traffic control office.