A Hybrid Features Based Detection Method for Inshore Ship Targets in SAR Imagery

Convolutional Neural Networks(CNNs)have recently attracted much attention in the ship detection from Synthetic Aperture Radar(SAR)images.However,compared with optical images,SAR ones are hard to understand.Moreover,due to the high similarity between the man-made targets near shore and inshore ships,the classical methods are unable to achieve effective detection of inshore ships.To mitigate the influence of onshore ship-like objects,this paper proposes an inshore ship detection method in SAR images by using hybrid features.Firstly,the sea-land segmentation is applied in the pre-processing to exclude obvious land regions from SAR images.Then,a CNN model is designed to extract deep features for identifying potential ship targets in both inshore and offshore water.On this basis,the high-energy point number of amplitude spectrum is further introduced as an important and delicate feature to suppress false alarms left.Finally,to verify the effectiveness of the proposed method,numerical and comparative studies are carried out in experiments on Sentinel-1 SAR images.

Automatic Feature Point Detection and Tracking of Human Actions in Time-of-flight Videos

Detecting feature points on the human body in video frames is a key step for tracking human movements. There have been methods developed that leverage models of human pose and classification of pixels of the body image. Yet, occlusion and robustness are still open challenges. In this paper, we present an automatic, model-free feature point detection and action tracking method using a time-of-flight camera. Our method automatically detects feature points for movement abstraction. To overcome errors caused by miss-detection and occlusion, a refinement method is devised that uses the trajectory of the feature points to correct the erroneous detections. Experiments were conducted using videos acquired with a Microsoft Kinect camera and a publicly available video set and comparisons were conducted with the state-of-the-art methods. The results demonstrated that our proposed method delivered improved and reliable performance with an average accuracy in the range of 90 %.The trajectorybased refinement also demonstrated satisfactory effectiveness that recovers the detection with a success rate of 93.7 %. Our method processed a frame in an average time of 71.1 ms.

Feature detection on point clouds via Gabriel Triangles creation and l1 normal reconstruction

In this paper, we present a robust subneighborhoods selection technique for feature detection on point clouds scattered over a piecewise smooth surface. The proposed method first identifies all potential features using covariance analysis of the local- neighborhoods. To further extract the accurate features from potential features, Gabriel triangles are created in local neighborhoods of each potential feature vertex. These triangles tightly attach to underlying surface and effectively reflect the local geometry struc- ture. Applying a shared nearest neighbor clustering algorithm on ~ 1 reconstructed normals of created triangle set, we classify the lo- cal neighborhoods of the potential feature vertex into multiple subneighborhoods. Each subneighborhood indicates a piecewise smooth surface. The final feature vertex is identified by checking whether it is locating on the intersection of the multiple surfaces. An advantage of this framework is that it is not only robust to noise, but also insensitive to the size of selected neighborhoods. Ex- perimental results on a variety of models are used to illustrate the effectiveness and robustness of our method.

Loop Closure Detection of Visual SLAM Based on Point and Line Features

For traditional loop closure detection algorithm,only using the vectorization of point features to build visual dictionary is likely to cause perceptual ambiguity.In addition,when scene lacks texture information,the number of point features extracted from it will be small and cannot describe the image effectively.Therefore,this paper proposes a loop closure detection algorithm which combines point and line features.To better recognize scenes with hybrid features,the building process of traditional dictionary tree is improved in the paper.The features with different flag bits were clustered separately to construct a mixed dictionary tree and word vectors that can represent the hybrid features,which can better describe structure and texture information of scene.To ensure that the similarity score between images is more reasonable,different similarity coefficients were set in different scenes,and the candidate frame with the highest similarity score was selected as the candidate closed loop.Experiments show that the point line comprehensive feature was superior to the single feature in the structured scene and the strong texture scene,the recall rate of the proposed algorithm was higher than the state of the art methods when the accuracy is 100%,and the algorithm can be applied to more diverse environments.

Face mask detection algorithm based on HSV+HOG features and SVM

To automatically detecting whether a person is wearing mask properly,we propose a face mask detection algorithm based on hue-saturation-value(HSV)+histogram of oriented gradient(HOG)features and support vector machines(SVM).Firstly,human face and five feature points are detected with RetinaFace face detection algorithm.The feature points are used to locate to mouth and nose region,and HSV+HOG features of this region are extracted and input to SVM for training to realize detection of wearing masks or not.Secondly,RetinaFace is used to locate to nasal tip area of face,and YCrCb elliptical skin tone model is used to detect the exposure of skin in the nasal tip area,and the optimal classification threshold can be found to determine whether the wear is properly according to experimental results.Experiments show that the accuracy of detecting whether mask is worn can reach 97.9%,and the accuracy of detecting whether mask is worn correctly can reach 87.55%,which verifies the feasibility of the algorithm.

General and robust voxel feature learning with Transformer for 3D object detection

The self-attention networks and Transformer have dominated machine translation and natural language processing fields,and shown great potential in image vision tasks such as image classification and object detection.Inspired by the great progress of Transformer,we propose a novel general and robust voxel feature encoder for 3D object detection based on the traditional Transformer.We first investigate the permutation invariance of sequence data of the self-attention and apply it to point cloud processing.Then we construct a voxel feature layer based on the self-attention to adaptively learn local and robust context of a voxel according to the spatial relationship and context information exchanging between all points within the voxel.Lastly,we construct a general voxel feature learning framework with the voxel feature layer as the core for 3D object detection.The voxel feature with Transformer(VFT)can be plugged into any other voxel-based 3D object detection framework easily,and serves as the backbone for voxel feature extractor.Experiments results on the KITTI dataset demonstrate that our method achieves the state-of-the-art performance on 3D object detection.

基于改进PointPillars的3D目标检测算法

针对自动驾驶场景下远距离及遮挡目标识别问题,对PointPillars算法进行了改进。引入了并行的空间注意力和通道注意力机制,增强了目标的位置信息及有用特征通道权重,提高了远距离目标的检测精度。在2D CNN骨干网络中引入了自适应空间特征融合模块,解决了特征拼接的信息丢失问题,提高了遮挡目标的检测精度。基于KITTI数据集在3种不同场景难度下分别对SECOND、PointPillars、改进PointPillars这3种算法进行了定量分析验证,并将改进的PointPillars算法进行可视化分析。定量分析表明,改进PointPillars算法在鸟瞰图模式下目标检测精度最大提升2.75%;在三维模式下目标检测精度最大提升2.93%;在AOS模式下目标检测精度最大提升4.05%,可视化结果表明改进的PointPillars算法能有效检测远距离及遮挡目标。

Detection of Multiscale Center Point Objects Based on Parallel Network

Anchor-based detectors are widely used in object detection.To improve the accuracy of object detection,multiple anchor boxes are intensively placed on the input image,yet.Most of which are invalid.Although the anchor-free method can reduce the number of useless anchor boxes,the invalid ones still occupy a high proportion.On this basis,this paper proposes a multiscale center point object detection method based on parallel network to further reduce the number of useless anchor boxes.This study adopts the parallel network architecture of hourglass-104 and darknet-53 of which the first one outputs heatmaps to generate the center point for object feature location on the output attribute feature map of darknet-53.Combining feature pyramid and CIoU loss function,this algorithm is trained and tested on MSCOCO dataset,increasing the detection rate of target location and the accuracy rate of small object detection.Though resembling the state-of-the-art two-stage detectors in overall object detection accuracy,this algorithm is superior in speed.

基于FVOIRGAN-Detection的车辆检测

为了解决点云处理过程中空间信息损失的问题,同时在融合过程中最大程度地提取可见光图像的纹理信息,本文提出了一种基于特征切片的激光点云与可见光图像融合车辆检测方法(FVOIRGAN-Detection)。在CrossGAN-Detection方法中加入了FVOI(Front View Based on Original Information)的点云处理思路,将点云投影到前视角度并把原始点云信息的各个维度切片为特征通道,在不降低网络性能的情况下显著提高点云信息利用效率。并且引入了相对概率的思想,采用鉴别器鉴别图像的相对真实概率替代绝对真实概率,使得融合图像提取的纹理信息更加接近真实的纹理信息。在KITTI数据集上进行检测性能实验验证结果表明,本文方法在容易、中等和困难三个类别中的AP指标分别达到97.67%、87.86%和79.03%。在光线受限的场景下,AP指标达到了88.49%,与CrossGAN-Detection方法相比提高了2.37%,提高了目标检测的性能。

Automatic Body Feature Extraction from Front and Side Images

Human body feature extraction based on 2D images provides an efficient method for many applications, e.g. non-contact body size measurements, constructing 3D human model and recognizing human actions. In this paper a systematic approach is proposed to detect feature points of human body automatically from its front and side images. Firstly, an efficient approach for silhouette and contour detection is used to represent the contour curves of a human body shape with Freeman’s 8-connected chain codes. The contour curves are considered as a number of segments connected together. Then, a series of feature points on human body are extracted based on the specified rules by measuring the differences between the directions of the segments. In total, 101 feature points with clearly geometric properties (that rather accurately reflect the bump or turning of the contours) are extracted automatically, including 27 points corresponding to the definitions of the landmarks about garment measurements. Finally, the proposed approach was tested on ten human subjects and the entire 101 feature points with specific geography geometrical characteristics were correctly extracted, indicating an effective and robust performance.

Medical equipments high precise detection technology basing on morphology-harris operator

Medical equipments related to life safety of human, it is important to detect by a high precise method. Image mosaic which based on Harris corner operator is a commonly used method in this area;Harris operator has low calculation burden, it is simple and stable, so it is more effective comparing with other feature point extracted operators. But in this algorithm, corner points can only be detected in a single-scale, there may be losing information of corner points, causing corner point location offset, extracting false corner points because of noise. In order to solve this question, the acquired images should be processed by dilation and erosion operation firstly, then do image mosaic. Results show that image noise can be eliminated effectively after those morphological processes, as well as the false positive noise generated by image glitch. The success rate of image mosaic and detection accuracy can be greatly improved through the Morphology-Harris operator. Measurement of precision instruments which based on this new method will improve the measurement accuracy, and the research in this area will promote the further development of machine vision technology.

远距离和遮挡下三维目标检测算法研究

针对现有三维目标检测算法对存在遮挡及距离较远目标检测效果差的问题,以基于点云的三维目标检测算法(3D object proposal generation and detection from point cloud,PointRCNN)为基础,对网络进行改进,提高三维目标检测精度。对区域生成网络(region proposal network,RPN)获取的提议区域(region of interest,ROI)体素化处理,同时构建不同尺度的区域金字塔来捕获更加广泛的兴趣点;加入点云Transformer模块来增强对网格中心点局部特征的学习;在网络中加入球查询半径预测模块,使得模型可以根据点云密度自适应调整球查询的范围。最后,对所提算法的有效性进行了试验验证,在KITTI数据集下对模型的性能进行评估测试,同时设计相应的消融试验验证模型中各模块的有效性。

图像语义特征引导与点云跨模态融合的三维目标检测方法

受到场景的复杂性和目标尺度变化、遮挡等影响,三维目标检测仍面临着诸多挑战.虽然跨模态特征融合图像和激光点云信息能够有效地提升三维目标检测性能,但在融合效果和检测性能上仍有待提升,为此,提出图像语义特征引导与点云跨模态融合的三维目标检测方法.首先设计图像语义特征学习网络,采用双分支自注意力并行计算方式,实现全局语义特征增强,降低目标错误分类;然后提出图像语义特征引导的局部融合模块,采用元素级数据拼接将检索的图像局部语义特征引导融合点云数据,更好地解决跨模态信息融合存在的语义对齐问题;提出多尺度再融合网络,设计融合特征与激光雷达点云交互模块,学习融合特征和不同分辨率特征间的再融合,提高网络的检测性能;最后采用4种任务损失实现anchor-free的三维目标检测.在KITTI和nuScenes数据集中与其他方法进行对比,针对三维目标检测准确率达87.15%,并且实验结果表明,文中方法优于对比方法,具有更优的三维检测性能.

结合轻量化YOLOv5s的动态视觉SLAM算法

针对视觉SLAM在真实环境中易受车辆、行人等运动物体影响而导致位姿估计降低的问题,提出一种结合轻量化YOLOv5s的动态视觉SLAM算法,将改进的轻量级YOLOv5s作为目标检测算法用于判断运动物体;结合提出的动态特征点剔除方法,剔除动态特征点,仅采用静态特征点进行位姿估计和地图跟踪。在TUM数据集进行实验,相较于ORB-SLAM3算法,改进后的算法在高动态序列上的位姿估计精度分别提升了89.29%、65.34%、94.42%,结果表明改进后的算法能够有效剔除动态特征点,提高了视觉SLAM算法在动态环境下的位姿估计精度和定位精度。

基于区域预推荐和特征富集的SOD R-CNN交通标志检测网络

基于区域的快速卷积神经网络存在资源的浪费和无法有效应对小目标检测的问题,提出基于高可能性区域推荐网络及特征富集的区域的小目标检测卷积神经网络架构。首先,采用区域推荐网络对锚点区域进行筛选,节约分类阶段的处理时间,提高了系统的处理速度。其次,为了解决无法有效检测小目标的问题,提出了融合视觉几何组16层网络的第三、第四、第五层特征信息的方法来强化特征表达的策略。最后,提出次要感兴趣区域的概念,将交通标志的上下文信息融合到目标特征表达中。这些策略提高了目标检测的准确率和速度。

一种基于目标检测的动态环境下视觉定位系统

传统的基于同时定位与建图模型的视觉定位方法需要满足目标点静止假设,但大多数小型机器人的实际应用场景为动态,这限制了现有视觉定位算法在小型机器人上的使用。为此,文中使用YOLOv5卷积神经网络对环境中的动态目标进行检测,然后剔除分布在图中的移动特征点,进而改进位姿估计准确性的动态消除方法,并将此方法集成于ORBSLAM2视觉定位系统。改进方案在TUM公共动态数据集上的测试结果表明,基于YOLOv5的检测方法能够快速、准确地识别场景中的动态目标,并显著降低动态环境下位姿估计的绝对误差和相对漂移,是一种有效的动态场景视觉定位方案。

城市道路场景下的被遮挡车辆检测算法研究

为了提高智能汽车在城市道路场景下对前方被遮挡车辆的检测精度,提出了一种双尺度点云密度扩展网络BPDE-Net,来解决因存在车辆被遮挡而导致的目标点云稀疏问题。第一阶段,将原始点云投影到语义分割后的图像标签上,并在遮挡区域内随机生成固定数目的虚拟点,采用混合插值法来关联虚拟点和实际投影点,得到的虚拟点再反向映射到点云空间;第二阶段,使用马氏距离来关联相邻体素间的点云,以此增加每个体素内的相似点云数量,通过改进注意力高斯矩阵来计算体素特征所对应的相对位置编码,用于关注通道内不同体素序列的相对位置。在KITTI数据集中选取了大量的车辆之间存在遮挡的城市道路场景进行对比试验,结果表明:BPDE-Net在3D视角和鸟瞰图视角下的被遮挡车辆平均检测精度(mAP)分别为79.2%和83.7%,相较于基线网络Second分别提高了11.2%和12.5%;点云密度增强模块和体素特征融合模块的mAP相较于目前主流的方法分别提高了3.9%与1.8%,提升了车辆在被遮挡场景下的可辨识度与鲁棒性。

融合体素图注意力的三维目标检测算法

目前基于点云的三维目标检测方法未能充分利用点云局部几何特征,导致对点云稀疏的目标检测效果不佳。为此,本文提出基于原始点云体素图注意力的两阶段三维目标检测算法(voxel graph attention region-CNN,VGT-RCNN)。通过多尺度体素特征插值计算网格中心点特征;在多尺度非空体素特征上构造局部图;通过图注意力机制对体素特征进行加权平均,充分提取并利用目标的局部几何特征完成检测。该算法主要针对当前二阶段算法在进行特征聚合时对不同体素特征的重要性考虑不足进行改进,引入可学习的权重矩阵,动态学习体素特性的权重,提高局部特征表达能力。本文在流行的KITTI自动驾驶数据集上进行了充分测试,取得了具有竞争力的检测效果,尤其是在对点云稀疏的汽车目标检测上,准确率有较大提高。本文还对检测效果进行了可视化分析。

基于消失点引导透视变换的车道线检测算法

目的为解决车道线的位置会随着车辆或相机的偏移发生变化而导致车道线检测准确率低和适应性差的问题,提出了一种基于消失点引导透视变换的车道线检测算法。方法首先,采用自适应消失点坐标引导更新透视变换矩阵,将车道图像转换为车道线保存完整的鸟瞰图;其次,将其颜色特征和边缘特征进行融合,得到精准的二值化图像;最后,根据直方图分析定位车道线的基点,采用滑动窗口搜索的方法提取候选的车道线像素,然后对搜索到的车道线像素进行多项式拟合。在不同的道路场景下测试算法的性能,并与其它同类算法进行对比分析。结果仿真结果表明,算法的准确率为94.12%,平均每帧耗时85.35ms,在检测精度和速度方面优于对比的算法。结论该算法能有效解决车道线位置的改变对车道线检测性能的影响,具有更高的准确率和较好的适应性,在阴影遮挡、车道破损、恶劣天气等复杂道路环境的检测下,表现出良好的鲁棒性。