Lightweight Cross-Modal Multispectral Pedestrian Detection Based on Spatial Reweighted Attention Mechanism

Multispectral pedestrian detection technology leverages infrared images to provide reliable information for visible light images, demonstrating significant advantages in low-light conditions and background occlusion scenarios. However, while continuously improving cross-modal feature extraction and fusion, ensuring the model’s detection speed is also a challenging issue. We have devised a deep learning network model for cross-modal pedestrian detection based on Resnet50, aiming to focus on more reliable features and enhance the model’s detection efficiency. This model employs a spatial attention mechanism to reweight the input visible light and infrared image data, enhancing the model’s focus on different spatial positions and sharing the weighted feature data across different modalities, thereby reducing the interference of multi-modal features. Subsequently, lightweight modules with depthwise separable convolution are incorporated to reduce the model’s parameter count and computational load through channel-wise and point-wise convolutions. The network model algorithm proposed in this paper was experimentally validated on the publicly available KAIST dataset and compared with other existing methods. The experimental results demonstrate that our approach achieves favorable performance in various complex environments, affirming the effectiveness of the multispectral pedestrian detection technology proposed in this paper.

An Expert System to Detect Political Arabic Articles Orientation Using CatBoost Classifier Boosted by Multi-Level Features

The number of blogs and other forms of opinionated online content has increased dramatically in recent years.Many fields,including academia and national security,place an emphasis on automated political article orientation detection.Political articles(especially in the Arab world)are different from other articles due to their subjectivity,in which the author’s beliefs and political affiliation might have a significant influence on a political article.With categories representing the main political ideologies,this problem may be thought of as a subset of the text categorization(classification).In general,the performance of machine learning models for text classification is sensitive to hyperparameter settings.Furthermore,the feature vector used to represent a document must capture,to some extent,the complex semantics of natural language.To this end,this paper presents an intelligent system to detect political Arabic article orientation that adapts the categorical boosting(CatBoost)method combined with a multi-level feature concept.Extracting features at multiple levels can enhance the model’s ability to discriminate between different classes or patterns.Each level may capture different aspects of the input data,contributing to a more comprehensive representation.CatBoost,a robust and efficient gradient-boosting algorithm,is utilized to effectively learn and predict the complex relationships between these features and the political orientation labels associated with the articles.A dataset of political Arabic texts collected from diverse sources,including postings and articles,is used to assess the suggested technique.Conservative,reform,and revolutionary are the three subcategories of these opinions.The results of this study demonstrate that compared to other frequently used machine learning models for text classification,the CatBoost method using multi-level features performs better with an accuracy of 98.14%.

Neural Network Based Multi-level Fuzzy Evaluation Model for Mechanical Kinematic Scheme

To implement a quantificational evaluation for mechanical kinematic scheme more effectively,a multi-level and multi-objective evaluation model is presented using neural network and fuzzy theory. Firstly,the structure of evaluation model is constructed according to evaluation indicator system. Then evaluation samples are generated and provided to train this model. Thus it can reflect the relation between attributive value and evaluation result,as well as the weight of evaluation indicator. Once evaluation indicators of each candidate are fuzzily quantified and fed into the trained network model,the corresponding evaluation result is outputted and the best alternative can be selected. Under this model,expert knowledge can be effectively acquired and expressed,and the quantificational evaluation can be implemented for kinematic scheme with multi-level evaluation indicator system. Several key problems on this model are discussed and an illustration has demonstrated that this model is feasible and can be regarded as a new idea for solving kinematic scheme evaluation.

Pedestrian attribute classification with multi-scale and multi-label convolutional neural networks

Pedestrian attribute classification from a pedestrian image captured in surveillance scenarios is challenging due to diverse clothing appearances,varied poses and different camera views. A multiscale and multi-label convolutional neural network( MSMLCNN) is proposed to predict multiple pedestrian attributes simultaneously. The pedestrian attribute classification problem is firstly transformed into a multi-label problem including multiple binary attributes needed to be classified. Then,the multi-label problem is solved by fully connecting all binary attributes to multi-scale features with logistic regression functions. Moreover,the multi-scale features are obtained by concatenating those featured maps produced from multiple pooling layers of the MSMLCNN at different scales. Extensive experiment results show that the proposed MSMLCNN outperforms state-of-the-art pedestrian attribute classification methods with a large margin.

Specific-Scene Oriented Pedestrian Detection in Visual Sensor Network

Pedestrian detection is one of the most important problems in the visual sensor network. Considering that the visual sensors have limited cap ability, we propose a pedestrian detection method with low energy consumption. Our method contains two parts: one is an Enhanced Self-Organizing Background Subtraction (ESOBS) based foreground segmentation module to obtain active areas in the observed region from the visual sensors; the other is an appearance model based detection module to detect the pedestrians from the foreground areas. Moreover, we create our own large pedestrian dataset according to the specific scene in the visual sensor network. Numerous experiments are conducted in both indoor and outdoor specific scenes. The experimental results show that our method is effective.

A New Method for Pedestrian Detection with Lightweight Backbone based on Yolov3 Network

The main purpose of YOLOv3,aiming to improve the detection speed and accuracy from current detection models,is to predict the center coordinates of(x,y)from the Bounding Box and its length,width through multiple layers of VGG Convolutional Neural Network(VGG-CNN)and uses the Darknet lightweight framework to process images at a faster speed.More specifically,our model has been reduced part of YOLOv3's complex and computationally intensive procedures and improved its algorithms to maintain the efficiency and accuracy of object detection.By this method,it performs a higher quality on mass object detection tasks with fewer detection errors.

A Real-Time Pedestrian Social Distancing Risk Alert System for COVID-19

The COVID-19 virus is usually spread by small droplets when talking,coughing and sneezing,so maintaining physical distance between people is necessary to slow the spread of the virus.The World Health Organization(WHO)recommends maintaining a social distance of at least six feet.In this paper,we developed a real-time pedestrian social distance risk alert system for COVID-19,whichmonitors the distance between people in real-time via video streaming and provides risk alerts to the person in charge,thus avoiding the problem of too close social distance between pedestrians in public places.We design a lightweight convolutional neural network architecture to detect the distance between people more accurately.In addition,due to the limitation of camera placement,the previous algorithm based on flat view is not applicable to the social distance calculation for cameras,so we designed and developed a perspective conversion module to reduce the image in the video to a bird’s eye view,which can avoid the error caused by the elevation view and thus provide accurate risk indication to the user.We selected images containing only person labels in theCOCO2017 dataset to train our networkmodel.The experimental results show that our network model achieves 82.3%detection accuracy and performs significantly better than other mainstream network architectures in the three metrics of Recall,Precision and mAP,proving the effectiveness of our system and the efficiency of our technology.

基于动作条件交互的高效行人过街意图预测

城市化的进程不断加速,人车冲突问题已成为现代社会亟待解决的重大难题。复杂交通场景下,行人横穿马路行为导致交通事故频发,准确、实时地预测行人过街意图对避免人车冲突、提高驾驶安全系数和保障行人安全至关重要。本文提出基于动作条件交互的高效行人过街意图预测框架(efficient action-conditioned interaction pedestrian crossing intention anticipation framework,EAIPF)来预测行人过街意图。EAIPF引入行人动作编码模块增强多模态动作模式下的表征能力,挖掘深层骨架上下文信息。同时,引入场景对象交互模块挖掘与对象交互信息,理解交通场景中高级语义线索。最后,意图预测模块融合行人动作特征和对象交互特征,实现行人过街意图的鲁棒预测。所提出的方法在两个公共数据集JAAD和PIE上验证算法性能,准确率分别达到了89%和90%,表明本文方法可以在复杂交通场景下准确预测行人穿越意图。

基于多尺度特征与互监督的拥挤行人检测

针对拥挤场景中,行人漏检率高、准确率低的问题,提出一种基于多尺度特征与互监督的拥挤行人检测网络。为了有效提取复杂场景中的行人特征信息,用PANet金字塔网络与混合空洞卷积相结合的网络提取特征信息。然后,设计了一种行人头部-全身互监督检测网络分别进行头部和全身检测,利用头部预测框和全身预测框的互监督获得更加准确的行人检测结果。所提出的网络在数据集CrowdHuman上取得了13.5%的MR^(-2)性能,相较于YOLOv5网络提升了3.6%,同时AP提升了3.5%;在CityPersons数据集上取得了48.2%的MR^(-2)性能,相较于YOLOv5网络提升了2.3%,同时AP提升了2.8%。实验结果表明,提出的网络在人群拥挤的密集场景中具有良好的检测效果。

基于多信息融合网络的行人轨迹预测方法

随着自动驾驶技术的不断发展,准确预测行人的未来轨迹已经成为确保系统安全和可靠的关键要素。然而,现有行人轨迹预测研究多数依赖于固定摄像头视角,进而限制了对行人运动的全面观测,因此难以直接应用于自动驾驶车辆自车视角(ego-vehicle)下的行人轨迹预测。针对该问题,本文提出了一种基于多行人信息融合网络(MPIFN)的自车视角行人轨迹预测方法。该方法通过融合社会信息、局部环境信息和行人时间信息,实现了对行人未来轨迹的准确预测。本文构建了一个局部环境信息提取模块,结合了可形变卷积与传统卷积和池化操作,旨在更有效地提取复杂环境中的局部信息。该模块通过动态调整卷积核的位置,增强了模型对不规则和复杂形状的适应能力。同时,构建了行人时空信息提取模块和多模态特征融合模块,以实现对社会信息与环境信息的充分融合。实验结果表明,该方法在JAAD和PSI两个自车视角下驾驶数据集上均取得了先进的性能。在JAAD数据集上,累积均方误差(CF_MSE)为4 063,累积平均均方误差(C_MSE)为829。在PSI数据集上平均相对偏差(ARB)和最终相对偏差(FRB)也分别在预测时间为0.5、1.0、1.5 s时取得了18.08、29.21、44.98和25.27、54.62、93.09的优异表现。

深度学习与图像融合的行人检测算法研究

为解决智能辅助驾驶技术中可见光摄像机受光照和气候影响而导致行人目标识别困难的问题。通过研究图像融合技术,结合深度卷积神经网络,实现并改进了一种道路行人目标检测算法。方法是利用多源传感器图像融合技术,采用可见光相机与红外热成像相机融合的策略,以Faster RCNN算法为基础,从改进网络结构、特征融合、优化模型训练等方面展开研究,对复杂环境下的行人检测与定位跟踪展开研究,提出一种基于图像融合技术和改进的深度卷积神经网络的道路行人目标检测算法。实验结果表明,该算法对复杂气候环境下行人目标检测提高了检测效率和准确率,增加了智能辅助驾驶汽车的安全性。

融合距离阈值和双向TCN的时空注意力行人轨迹预测模型

为解决因缺乏部分行人建模思想、缺少时间维度的全局视野和忽略行人交互模式多样性,而导致交互建模不充分、低预测精度等问题,提出基于Social-STGCNN(social spatio-temporal graph convolutional neural network)的改进模型STG-DTBTA(spatio-temporal graph distance threshold Bi-TCN attention)。首先,构建PPM(partial pedestrian module)模块,对不满足距离阈值等约束条件的行人交互连接剪枝以去噪。其次,引入时空注意力机制,空间注意力动态分配交互权重,并设置多个注意力头以处理交互多样性问题;时间注意力捕捉时序数据的时间依赖关系。最后,采用双向TCN增加全局视野以捕捉轨迹数据中的动态模式和趋势,并采用门控机制融合双向特征。在ETH和UCY数据集上的实验结果表明,与Social-STGCNN相比,STG-DTBTA在维持参数量与推理时间接近的情况下,ADE平均降低8%,FDE平均降低16%。STG-DTBTA具有良好的交互建模能力、模型性能和预测效果。

促进时空可达的高校新校区步行路网优化——以重庆医科大学缙云校区为例

可达性是高校校园步行出行的基本要求.然而,我国高校新校区既有步行路网存在局部空间可达水平偏低、近距离绕行等问题.本研究以重庆医科大学缙云校区为例,采用空间句法理论下的线段模型和ArcGIS最优路径分析模型,量化其步行时空可达水平;结合步行空间实际使用水平,提出校园步行路网的改善方案并评估方案的步行可达性.研究发现:(1)分级的步行路网可提升全局和局部空间可达性,精细的道路设施设计可确保实际使用的可达性;(2)有序的游憩步道可增强空间感知的可达性;(3)以吸引点为导向的步行捷径可改善时间可达性.

面向交通场景的轻量级行人检测算法

针对交通场景下行人检测模型网络复杂、参数量大以及难以在低性能设备上部署的问题,基于YOLOv5s网络模型提出了一种改进的轻量级行人检测算法。首先,使用Ghost模块重构YOLOv5s网络进行特征提取,降低模型的参数量和计算量,提高推理速度。其次,引入坐标注意力机制提高模型对目标特征的提取能力,提升其对小目标行人的检测效果。最后,采用SIoU损失函数加快模型的收敛速度,提高模型的识别准确率。实验结果表明,改进后的算法能保证较高的检测精度,与原始YOLOv5s算法相比参数量减少47.1%,计算量减少48.7%,提高了交通场景下行人检测的速度且易于部署。

基于信息分形的行人轨迹预测方法

行人轨迹预测应用十分广泛,比如自动驾驶、机器人导航等。在轨迹预测中,一些不确定信息给轨迹预测任务带来了挑战,比如判别器中对轨迹信息判别的不确定,复杂的交互信息。在不确定信息处理科学领域,信息分形能有效处理不确定信息的不确定性和复杂性。受此启发,为了充分处理判别器中轨迹信息判别的不确定性,提升预测精度,该文提出了基于信息分形的轨迹预测方法。首先,场景信息和历史轨迹信息被特征提取模块提取。然后,通过注意力模块获取到场景-行人之间的交互信息与行人-行人之间的交互信息。最后基于生成对抗网络和信息分形生成合理的轨迹。在两个公共数据集ETH/UCY上实验表明,该方法能有效处理轨迹信息的不确定性,提高轨迹预测的精度。比如突然转弯、从后方超越前人、避让等行为的轨迹都能有效预测。在平均位移误差(ADE)和终点位移误差(FDE)上相比基准模型误差平均降低了11.11%和23.48%。

高密度立体复合街区步行系统构建--以郑州东站东广场核心区为例

高密度立体复合街区逐渐成为城市空间组成的核心,构建立体步行系统能够满足不同空间行人多样化和高品质出行需求。通过总结高密度立体复合街区特征和立体步行系统的组成与功能,明确步行系统与街区空间形态关系。提出高密度立体复合街区内步行系统构建要点,包括深入调查研究、精准预测需求、合理布局网络和明确落地实施,并进一步提出创新的关键点是缓解高密度街区对行人的压迫感、满足复合街区内多元化的步行需求和统筹街区空间的立体化。以郑州东站东广场核心区为例,分析其集客运枢纽、商务、商业、休闲、居住等多功能复合的空间特征和指标规模,以及立体复杂的地下空间综合开发形式,提出坚持以人为本,打造多层次步行空间,构建地下、地面、地上多维度立体步行网络,且以灵活便捷的空中连廊建设形式和面向实施的指标控制要求保障方案落地实施,以支撑街区内绿色低碳、健康可持续发展。

基于人体运动识别约束的室内定位方法

针对传统的行人航迹推算(Pedestrian Dead Reckoning,PDR)方法无法满足多运动状态下的定位问题,提出了一种基于神经网络运动识别辅助室内定位的方法。构建出卷积神经网络(Convolutional Neural Network,CNN)和门控循环单元(Gated Recurrent Unit,GRU)组合的神经网络模型,用于识别人体的运动状态并完成分类。根据运动分类的结果应用到行人航迹推算中,分析和筛选运动参数特征作为算法的阈值约束条件来提高定位精度。在算法中运动步数由合加速度计数据波形检测得到,步长由运动状态的特征自适应调整步长模型。通过实验验证,CNN-GRU模型在自建数据集上的准确率达到99.6%。将识别结果应用到PDR中,在112 m 4种动作的矩形路线中定位误差为1.8 m,误差远低于传统PDR定位的19.9 m。实验结果验证了该方法的可行性。

Cycle GAN-MF:A Cycle-consistent Generative Adversarial Network Based on Multifeature Fusion for Pedestrian Re-recognition

In pedestrian re-recognition,the traditional pedestrian re-recognition method will be affected by the changes of background,veil,clothing and so on,which will make the recognition effect decline.In order to reduce the impact of background,veil,clothing and other changes on the recognition effect,this paper proposes a pedestrian re-recognition method based on the cycle-consistent generative adversarial network and multifeature fusion.By comparing the measured distance between two pedestrians,pedestrian re-recognition is accomplished.Firstly,this paper uses Cycle GAN to transform and expand the data set,so as to reduce the influence of pedestrian posture changes as much as possible.The method consists of two branches:global feature extraction and local feature extraction.Then the global feature and local feature are fused.The fused features are used for comparison measurement learning,and the similarity scores are calculated to sort the samples.A large number of experimental results on large data sets CUHK03 and VIPER show that this new method reduces the influence of background,veil,clothing and other changes on the recognition effect.

面向复杂光照的舞台演员检测

复杂舞台场景存在多个光源产生的偏色和光照不均匀问题,严重影响了演员检测的精度.针对上述问题,提出一种基于伪多模态融合的演员检测方法.首先随机选取一种光照处理方法构建增强图像,与原图像构成伪多模态图像对;然后在增强图像中以演员关键点建立候选集合,从集合中随机选取部分关键点所在的区域构建增强补丁集合,并将补丁替换到原始图像中进行训练;最后在传统特征金字塔网络的基础上借鉴Transformer编码器的构建形式,利用视觉注意力模块构建视觉注意力编码器,强化多尺度特征的交互逻辑.在自建4543幅包含舞台演员的图像数据集上与3个模型进行组合,舞台演员检测的均值平均精度分别提升0.4%~2.9%,表明所提方法能够较好地降低偏色和不均匀光照的影响.

面向拥挤行人检测的改进YOLOv7算法

针对拥挤行人检测场景下检测算法容易产生漏检与误检的问题,提出一种改进的YOLOv7拥挤行人检测算法。在骨干网络中引入BiFormer视觉变换器和改进的高效层聚合网络(RC-ELAN)模块,通过自注意力机制与注意力模块使骨干网络更多聚焦于被遮挡行人的重要特征,有效缓解了目标特征缺失对检测造成的负面影响。采用基于双向特征金字塔网络思想的改进颈部网络,通过转置卷积和改进的Rep-ELAN-W模块使模型可以高效利用中低维特征图中的小目标特征信息,有效提升了模型的小目标行人检测性能。引入高效的完全交并比损失函数,使模型可以进一步收敛至更高精度。在含有大量小目标遮挡行人的WiderPerson数据集上的实验结果表明,与YOLOv7、YOLOv5、YOLOX算法相比,改进的YOLOv7算法的交并比阈值分别取0.5和0.5~0.95时的平均精准度提升了2.5和2.8、9.9和7.1、12.3和10.7个百分点,可较好地应用于拥挤行人检测场景。