Depth-Guided Vision Transformer With Normalizing Flows for Monocular 3D Object Detection

Monocular 3D object detection is challenging due to the lack of accurate depth information.Some methods estimate the pixel-wise depth maps from off-the-shelf depth estimators and then use them as an additional input to augment the RGB images.Depth-based methods attempt to convert estimated depth maps to pseudo-LiDAR and then use LiDAR-based object detectors or focus on the perspective of image and depth fusion learning.However,they demonstrate limited performance and efficiency as a result of depth inaccuracy and complex fusion mode with convolutions.Different from these approaches,our proposed depth-guided vision transformer with a normalizing flows(NF-DVT)network uses normalizing flows to build priors in depth maps to achieve more accurate depth information.Then we develop a novel Swin-Transformer-based backbone with a fusion module to process RGB image patches and depth map patches with two separate branches and fuse them using cross-attention to exchange information with each other.Furthermore,with the help of pixel-wise relative depth values in depth maps,we develop new relative position embeddings in the cross-attention mechanism to capture more accurate sequence ordering of input tokens.Our method is the first Swin-Transformer-based backbone architecture for monocular 3D object detection.The experimental results on the KITTI and the challenging Waymo Open datasets show the effectiveness of our proposed method and superior performance over previous counterparts.

Design of a road vehicle detection system based on monocular vision

In order to decrease vehicle crashes, a new rear view vehicle detection system based on monocular vision is designed. First, a small and flexible hardware platform based on a DM642 digtal signal processor （DSP） micro-controller is built. Then, a two-step vehicle detection algorithm is proposed. In the first step, a fast vehicle edge and symmetry fusion algorithm is used and a low threshold is set so that all the possible vehicles have a nearly 100% detection rate （TP） and the non-vehicles have a high false detection rate （FP）, i. e., all the possible vehicles can be obtained. In the second step, a classifier using a probabilistic neural network （PNN） which is based on multiple scales and an orientation Gabor feature is trained to classify the possible vehicles and eliminate the false detected vehicles from the candidate vehicles generated in the first step. Experimental results demonstrate that the proposed system maintains a high detection rate and a low false detection rate under different road, weather and lighting conditions.

Mobile Robot Localization and Navigation System Based on Monocular Vision

A system for mobile robot localization and navigation was presented.With the proposed system,the robot can be located and navigated by a single landmark in a single image.And the navigation mode may be following-track,teaching and playback,or programming.The basic idea is that the system computes the differences between the expected and the recognized position at each time and then controls the robot in a direction to reduce those differences.To minimize the robot sensor equipment,only one omnidirectional camera was used.Experiments in disturbing environments show that the presented algorithm is robust and easy to implement,without camera rectification.The rootmean-square error(RMSE) of localization is 1.4,cm,and the navigation error in teaching and playback is within 10,cm.

Autonomous Landing of Small Unmanned Aerial Rotorcraft Based on Monocular Vision in GPS-denied Area

Focusing on the low-precision attitude of a current small unmanned aerial rotorcraft at the landing stage, the present paper proposes a new attitude control method for the GPS-denied scenario based on the monocular vision. Primarily, a robust landmark detection technique is developed which leverages the well-documented merits of supporting vector machines (SVMs) to enable landmark detection. Then an algorithm of nonlinear optimization based on Newton iteration method for the attitude and position of camera is put forward to reduce the projection error and get an optimized solution. By introducing the wavelet analysis into the adaptive Kalman filter, the high frequency noise of vision is filtered out successfully. At last, automatic landing tests are performed to verify the method's feasibility and effectiveness. © 2014 Chinese Association of Automation.

Mobile Robot Hierarchical Simultaneous Localization and Mapping Using Monocular Vision

A hierarchical mobile robot simultaneous localization and mapping (SLAM) method that allows us to obtain accurate maps was presented. The local map level is composed of a set of local metric feature maps that are guaranteed to be statistically independent. The global level is a topological graph whose arcs are labeled with the relative location between local maps. An estimation of these relative locations is maintained with local map alignment algorithm, and more accurate estimation is calculated through a global minimization procedure using the loop closure constraint. The local map is built with Rao-Blackwellised particle filter (RBPF), where the particle filter is used to extending the path posterior by sampling new poses. The landmark position estimation and update is implemented through extended Kalman filter (EKF). Monocular vision mounted on the robot tracks the 3D natural point landmarks, which are structured with matching scale invariant feature transform (SIFT) feature pairs. The matching for multi-dimension SIFT features is implemented with a KD-tree in the time cost of O(lbN). Experiment results on Pioneer mobile robot in a real indoor environment show the superior performance of our proposed method.

Calibration of laser beam direction based on monocular vision

In the laser displacement sensors measurement system,the laser beam direction is an important parameter.Particularly,the azimuth and pitch angles are the most important parameters to a laser beam.In this paper,based on monocular vision,a laser beam direction measurement method is proposed.First,place the charge coupled device(CCD)camera above the base plane,and adjust and fix the camera position so that the optical axis is nearly perpendicular to the base plane.The monocular vision localization model is established by using circular aperture calibration board.Then the laser beam generating device is placed and maintained on the base plane at fixed position.At the same time a special target block is placed on the base plane so that the laser beam can project to the special target and form a laser spot.The CCD camera placed above the base plane can acquire the laser spot and the image of the target block clearly,so the two-dimensional(2D)image coordinate of the centroid of the laser spot can be extracted by correlation algorithm.The target is moved at an equal distance along the laser beam direction,and the spots and target images of each moving under the current position are collected by the CCD camera.By using the relevant transformation formula and combining the intrinsic parameters of the target block,the2D coordinates of the gravity center of the spot are converted to the three-dimensional(3D)coordinate in the base plane.Because of the moving of the target,the3D coordinates of the gravity center of the laser spot at different positions are obtained,and these3D coordinates are synthesized into a space straight line to represent the laser beam to be measured.In the experiment,the target parameters are measured by high-precision instruments,and the calibration parameters of the camera are calibrated by a high-precision calibration board to establish the corresponding positioning model.The measurement accuracy is mainly guaranteed by the monocular vision positioning accuracy and the gravity center extraction accuracy.The experimental results show the maximum error of the angle between laser beams reaches to0.04°and the maximum error of beam pitch angle reaches to0.02°.

Research on Vehicle Anti-collision Technique Based on Monocular Vision

Vehicle anti-collision technique is a hot topic in the research area of Intelligent Transport System. The research on preceding vehicles detection and the distance measurement, which are the key techniques, makes great contributions to safe-driving. This paper presents a method which can be used to detect preceding vehicles and get the distance between own car and the car ahead. Firstly, an adaptive threshold method is used to get shadow feature, and a shadow!area merging approach is used to deal with the distortion of the shadow border. Region of interest(ROI) is obtained using shadow feature. Then in the ROI, symmetry feature is analyzed to verify whether there are vehicles and to locate the vehicles. Finally, using monocular vision distance measurement based on camera interior parameters and geometrical reasoning, we get the distance between own car and the preceding one. Experimental results show that the proposed method can detect the preceding vehicle effectively and get the distance between vehicles accurately.

AB012. The effects of monocular deprivation do not accumulate across days in adults with normal vision

Background:We investigate whether changes in visual plasticity induced by monocular deprivation can be maintained across multiple days.It has been known that monocular deprivation strengthens the deprived eye in adults with normal vision for a short period of time(30-60 minutes).This has been shown through a variety of visual tasks such as binocular combination and rivalry.Methods:Ten subjects were recruited and patched for five consecutive days for two hours.We used a binocular phase combination task to measure the subjects’sensory eye balances.We initially measured their baseline of sensory eye balance,patched their dominant eye,and then conducted post-patching measurements at 0,3,6,12,24 and 48 minutes after patching.Results:We performed a 2-way ANOVA(Before vs.after patching×Day);we found that although the effect of monocular deprivation on the deprived eye was significant,F(1,9)=17.32,P=0.002,the effect of Day was not.Conclusions:Hence we found no accumulation of the patching effect across five days in healthy adults.This suggests that the degree of remnant neural plasticity in adult primary visual cortex may be too limited to be exploited therapeutically.

基于Vision Transformer的虹膜——人脸多特征融合识别研究

为了提高生物特征识别系统的准确性和鲁棒性,本文研究基于计算机视觉的虹膜—人脸多特征融合识别方法。本文对面部图像中虹膜区域进行提取以及预处理,采用对比度增强和归一化操作,加强了特征提取的一致性,提升了图像质量。为了获取丰富的深度特征,本文使用Vision Transformer模型对预处理后的虹膜和面部图像进行特征提取。利用多头注意力机制将虹膜和面部的多模态特征信息进行融合,再利用全连接层进行分类识别。试验结果表明,该方法识别性能优秀,识别准确性显著提升。

基于Vision Transformer的小麦病害图像识别算法

小麦白粉病、赤霉病和锈病是危害小麦产量的三大病害。为提高小麦病害图像的识别准确率,构建一种基于Vision Transformer的小麦病害图像识别算法。首先,通过田间拍摄的方式收集包含小麦白粉病、赤霉病和锈病3种病害在内的小麦病害图像,并对原始图像进行预处理,建立小麦病害图像识别数据集;然后,基于改进的Vision Transformer构建小麦病害图像识别算法,分析不同迁移学习方式和数据增强对模型识别效果的影响。试验可知,全参数迁移学习和数据增强能明显提高Vision Transformer模型的收敛速度和识别精度。最后,在相同时间条件下,对比Vision Transformer、AlexNet和VGG16算法在相同数据集上的表现。试验结果表明,Vision Transformer模型对3种小麦病害图像的平均识别准确率为96.81%,相较于AlexNet和VGG16模型识别准确率分别提高6.68%和4.94%。

基于Vision Transformer与迁移学习的裤装廓形识别与分类

针对裤装廓形识别与分类模型的分类不准确问题,文章采用带有自注意力机制的Vision Transformer模型实现裤装廓形图像的分类,对于图片背景等无关信息对廓形识别的干扰,添加自注意力机制,增强有用特征通道。为防止因裤型样本数据集较少产生过拟合问题,可通过迁移学习方法对阔腿裤、喇叭裤、紧身裤、哈伦裤4种裤装廓形进行训练和验证,将改进的Vision Transformer模型与传统CNN模型进行对比实验,验证模型效果。实验结果表明:使用Vision Transformer模型在4种裤装廓形分类上的分类准确率达到97.72%,与ResNet-50和MobileNetV2模型相比均有提升,可为服装廓形的图像分类识别提供有力支撑,在实际服装领域中有较高的使用价值。

细粒度图像分类上Vision Transformer的发展综述

细粒度图像分类(fine-grained image classification,FGIC)一直是计算机视觉领域中的重要问题。与传统图像分类任务相比,FGIC的挑战在于类间对象极其相似,使任务难度进一步增加。随着深度学习的发展,Vision Transformer(ViT)模型在视觉领域掀起热潮,并被引入到FGIC任务中。介绍了FGIC任务所面临的挑战,分析了ViT模型及其特性。主要根据模型结构全面综述了基于ViT的FGIC算法,包括特征提取、特征关系构建、特征注意和特征增强四方面内容,对每种算法进行了总结,并分析了它们的优缺点。通过对不同ViT模型在相同公用数据集上进行模型性能比较,以验证它们在FGIC任务上的有效性。最后指出了目前研究的不足,并提出未来研究方向,以进一步探索ViT在FGIC中的潜力。

Collaborative positioning for swarms:A brief survey of vision,LiDAR and wireless sensors based methods

As positioning sensors,edge computation power,and communication technologies continue to develop,a moving agent can now sense its surroundings and communicate with other agents.By receiving spatial information from both its environment and other agents,an agent can use various methods and sensor types to localize itself.With its high flexibility and robustness,collaborative positioning has become a widely used method in both military and civilian applications.This paper introduces the basic fundamental concepts and applications of collaborative positioning,and reviews recent progress in the field based on camera,LiDAR(Light Detection and Ranging),wireless sensor,and their integration.The paper compares the current methods with respect to their sensor type,summarizes their main paradigms,and analyzes their evaluation experiments.Finally,the paper discusses the main challenges and open issues that require further research.

基于Vision Transformer的阿尔茨海默病分类研究

为了有效地提升对阿尔茨海默病(AD)的磁共振成像(MRI)图像分类准确率,提出一种LC(Layer-Cut)-ViT方法。该方法通过引入Vision Transformer(ViT)的自注意力机制对MRI图像进行层切分,使模型能更好地理解图像的全局信息,同时突出切片间的特征关系。此外,通过配准、颅骨分离算法提取MRI图像的脑部组织部分,进一步提升模型的性能。实验结果显示,所提方法对阿尔茨海默病的MRI图像具有较好的分类能力。

Frequency and associated factors of accommodation and non-strabismic binocular vision dysfunction among medical university students

AIM:To investigate the frequency and associated factors of accommodation and non-strabismic binocular vision dysfunction among medical university students.METHODS:Totally 158 student volunteers underwent routine vision examination in the optometry clinic of Guangxi Medical University.Their data were used to identify the different types of accommodation and nonstrabismic binocular vision dysfunction and to determine their frequency.Correlation analysis and logistic regression were used to examine the factors associated with these abnormalities.RESULTS:The results showed that 36.71%of the subjects had accommodation and non-strabismic binocular vision issues,with 8.86%being attributed to accommodation dysfunction and 27.85%to binocular abnormalities.Convergence insufficiency(CI)was the most common abnormality,accounting for 13.29%.Those with these abnormalities experienced higher levels of eyestrain(χ2=69.518,P<0.001).The linear correlations were observed between the difference of binocular spherical equivalent(SE)and the index of horizontal esotropia at a distance(r=0.231,P=0.004)and the asthenopia survey scale(ASS)score(r=0.346,P<0.001).Furthermore,the right eye's SE was inversely correlated with the convergence of positive and negative fusion images at close range(r=-0.321,P<0.001),the convergence of negative fusion images at close range(r=-0.294,P<0.001),the vergence facility(VF;r=-0.234,P=0.003),and the set of negative fusion images at far range(r=-0.237,P=0.003).Logistic regression analysis indicated that gender,age,and the difference in right and binocular SE did not influence the emergence of these abnormalities.CONCLUSION:Binocular vision abnormalities are more prevalent than accommodation dysfunction,with CI being the most frequent type.Greater binocular refractive disparity leads to more severe eyestrain symptoms.

基于Vision Transformer和迁移学习的家庭领域哭声识别

针对SVM等传统机器学习算法准确率低和当前使用CNN处理家庭领域哭声识别在不同婴儿间出现泛化能力差的问题,提出了一种基于Vision Transformer和迁移学习的婴儿哭声音频分类算法。首先,为实现数据集样本的扩增,采用了包括梅尔频谱转换和数据增强的数据预处理技术,进而达到了增强模型鲁棒性的目的。而后,在微调后的Vision Transformer模型上进行迁移学习训练,同时,训练过程中利用了LookAhead优化器来不断调整模型参数以避免过拟合,最终实验实现了对婴儿哭声音频的自动分类。实验结果表明,本实验模型相比其他深度学习模型具有更高的精确率和更快的收敛速度,同时还能有效地学习到婴儿哭声中更具区分性的特征。可以在新生儿监护、听力筛查和异常检测等领域中发挥重要作用。

基于Vision Transformer和迁移学习的垃圾图像分类研究

为解决垃圾图像分类中分类准确率低及小样本类别性能差的问题,以生活垃圾图像为研究对象,以正确识别生活垃圾类别为研究目标,利用Vision Transformer模型为分类网络架构,使用迁移学习机制实现该模型在华为云垃圾分类数据集上的训练及分类推理。实验结果表明,基于注意力机制的分类模型相较于基于卷积结构的ResNet、DenseNet分类模型具有更高的分类准确率,可达96%,同时测试集的混淆矩阵表明Vision Transformer分类模型在样本不均衡数据集中对于小样本类别也具有较高的准确率,具有实际部署、推理的应用价值。

基于改进Vision Transformer的道岔故障智能诊断

道岔故障种类繁多,特征复杂,存在检测难、分类难等问题,导致故障排查效率低下,对铁路运输安全构成威胁。Vision Transformer模型在图像分类方面具有较高准确度,但是其处理的是图像块,而不是传统的像素级特征,在某些情况下可能会影响曲线局部信息的获取。针对上述情况,提出一种基于改进Vision Transformer模型的故障曲线分类算法。首先,对典型道岔故障及原因进行梳理分类,指出几种典型的道岔故障;其次,对使用道岔动作电流数据生成的图像尺寸进行调整并根据故障图像特点进行数据增强,使用ResNet网络取代原Vision Transformer模型中的故障图像分块机制进行特征提取,同时采用相对位置编码增强模型的适应性和泛化能力;最后,利用模型的多头自注意力机制,综合全局与局部信息进行分类,并得到分类权重。经过实验验证,本文道岔故障分类识别总体准确率达99.77%,各分类识别的平均精确率达99.78%,与原模型相比,在训练集和验证集上的识别精度分别提升了5.4%和2.4%。为了更好地理解模型的性能,采用Grad-CAM方法将迭代过程可视化,剖析了模型关注区域的变化过程,并在测试集上与VGG-16、DenseNet121等经典分类模型进行性能对比;通过ROC曲线评估分类效果,显示改进的模型取得更优结果。研究结果为道岔故障识别分类提供了新的理论支持,并为未来的研究提供了新的思路和方法。

基于Vision Transformer-LSTM(ViTL)的多时序遥感影像农作物分类方法

针对当前遥感农作物分类研究中深度学习模型对光谱时间和空间信息特征采样不足,农作物提取仍然存在边界模糊、漏提、误提的问题,提出了一种名为视觉Transformer-长短期记忆递归神经网络(Vision Transformer-long short term memory,ViTL)的深度学习模型,ViTL模型集成了双路Vision-Transformer特征提取、时空特征融合和长短期记忆递归神经网络(LSTM)时序分类等3个关键模块,双路Vision-Transformer特征提取模块用于捕获图像的时空特征相关性,一路提取空间分类特征,一路提取时间变化特征;时空特征融合模块用于将多时特征信息进行交叉融合;LSTM时序分类模块捕捉多时序的依赖关系并进行输出分类。综合利用基于多时序卫星影像的遥感技术理论和方法,对黑龙江省齐齐哈尔市讷河市作物信息进行提取,研究结果表明,ViTL模型表现出色,其总体准确率(Overall Accuracy,OA)、平均交并比(Mean Intersection over Union,MIoU)和F1分数分别达到0.8676、0.6987和0.8175,与其他广泛使用的深度学习方法相比,包括三维卷积神经网络(3-D CNN)、二维卷积神经网络(2-D CNN)和长短期记忆递归神经网络(LSTM),ViTL模型的F1分数提高了9%~12%,显示出显著的优越性。ViTL模型克服了面对多时序遥感影像的农作物分类任务中的时间和空间信息特征采样不足问题,为准确、高效地农作物分类提供了新思路。

Artificial hawk-eye camera for foveated, tetrachromatic, and dynamic vision

With the rapid development of drones and autonomous vehicles, miniaturized and lightweight vision sensors that can track targets are of great interests. Limited by the flat structure, conventional image sensors apply a large number of lenses to achieve corresponding functions, increasing the overall volume and weight of the system.