主动迁移学习的海上任意方向船只目标检测

在基于深度学习的遥感图像目标检测任务中,船只目标通常呈现出任意方向排列的特性,而常见的水平框目标检测算法一般不能满足此类场景的应用需求。因此本文在单阶段Anchor-Free目标检测器CenterNet的基础上加入旋转角度预测分支,使其能输出旋转边界框,以用于海上船只目标的检测。同时针对海上船只遥感数据集仅有水平边界框标注,无法直接适用于旋转框目标检测,且人工手动标注旋转框标签成本较高的问题,提出一种主动迁移学习的旋转框标签生成方法。首先,提出一种水平框-旋转框约束筛选算法,通过水平真值边界框来对旋转预测框进行监督约束,筛选出检测精度较高的图像加入训练集,然后通过迭代这一过程筛选出更多的图像,最后通过标签类别匹配,完成对数据集的旋转框自动化标注工作。本文最终对海上船只遥感图像数据集BDCI中约65.59%的图片进行旋转框标注,并手动标注部分未标注的图片作为测试集,将本文方法标注的图片作为训练集进行验证,评估指标AP50达到90.41%,高于其他旋转框检测器,从而表明本文方法的有效性。     

使用贝叶斯优化对遥感影像目标进行精确定位

针对大尺寸遥感影像目标检测中检测边框不精确的问题,提出使用高斯过程贝叶斯优化对遥感影像中的目标进行精确检测与定位。研究分为两个阶段,第一阶段使用基于边缘信息的EdgeBoxes算法对大尺寸遥感影像进行目标候选区域的选取,用分类器得到初始检测结果;为了得到更加准确的边框,在第二阶段,基于高斯过程的贝叶斯优化对每个目标的边框进行微调：①以目标初始边框为基准,在其周围选取与初始边框相交的边框集合,并得到一个高斯过程分布;②使用贝叶斯优化估计出下一个边框,并将其加入边框集;③求分类器对所有边框的得分,得分最高的边框作为下次迭代的基准边框;④重复若干次贝叶斯优化后得到最终的边框。实验结果表明：EdgeBoxes方法以较少的候选框可以得到较大的召回率,使用高斯过程的贝叶斯优化可以明显地提高检测边框的精度。     

局部双目视差回归的目标距离估计

目的 双目视觉是目标距离估计问题的一个很好的解决方案。现有的双目目标距离估计方法存在估计精度较低或数据准备较繁琐的问题,为此需要一个可以兼顾精度和数据准备便利性的双目目标距离估计算法。方法 提出一个基于R-CNN（region convolutional neural network）结构的网络,该网络可以实现同时进行目标检测与目标距离估计。双目图像输入网络后,通过主干网络提取特征,通过双目候选框提取网络以同时得到左右图像中相同目标的包围框,将成对的目标框内的局部特征输入目标视差估计分支以估计目标的距离。为了同时得到左右图像中相同目标的包围框,使用双目候选框提取网络代替原有的候选框提取网络,并提出了双目包围框分支以同时进行双目包围框的回归;为了提升视差估计的精度,借鉴双目视差图估计网络的结构,提出了一个基于组相关和3维卷积的视差估计分支。结果 在KITTI（Karlsruhe Institute of Technology and Toyota Technological Institute）数据集上进行验证实验,与同类算法比较,本文算法平均相对误差值约为3.2%,远小于基于双目视差图估计算法（11.3%）,与基于3维目标检测的算法接近（约为3.9%）。另外,提出的视差估计分支改进对精度有明显的提升效果,平均相对误差值从5.1%下降到3.2%。通过在另外采集并标注的行人监控数据集上进行类似实验,实验结果平均相对误差值约为4.6%,表明本文方法可以有效应用于监控场景。结论 提出的双目目标距离估计网络结合了目标检测与双目视差估计的优势,具有较高的精度。该网络可以有效运用于车载相机及监控场景,并有希望运用于其他安装有双目相机的场景。     

提升预测框定位稳定性的视频目标检测

目的 目前视频目标检测（object detection from video）领域大量研究集中在提升预测框定位准确性,对于定位稳定性提升的研究则较少。然而,预测框定位稳定性对多目标跟踪、车辆行驶控制等算法具有重要影响,为提高预测框定位稳定性,本文提出了一种扩张性非极大值抑制（expanded non-maximum suppression,Exp_NMS）方法和帧间平滑策略（frame bounding box smooth,FBBS）。方法 目标检测阶段使用YOLO（you only look once）v3神经网络,非极大值抑制阶段通过融合多个预测框信息得出结果,增强预测框在连续视频流中的稳定性。后续利用视频相邻帧信息关联的特点,对预测框进行平滑处理,进一步提高预测框定位稳定性。结果 选用UA-DETRAC（University at Albany detection and tracking benchmark dataset）数据集进行分析实验,使用卡尔曼滤波多目标跟踪算法进行辅助验证。本文在MOT（multiple object tracking）评价指标基础上,设计了平均轨迹曲折度（average track-tortuosity,AT）来直观、量化地衡量预测框定位稳定性及跟踪轨迹的平滑度。实验结果表明,本文方法几乎不影响预测框定位准确性,且对定位稳定性有大幅改善,相应跟踪质量得到显著提升。测试视频的MOTA（multiple object tracking accuracy）提升6.0%、IDs（identity switches）减少16.8%,跟踪FP（false positives）类型错误下降45.83%,AT下降36.57%,mAP（mean average precision）仅下降0.07%。结论 从非极大值抑制和前后帧信息关联两个角度设计相关策略,经实验验证,本文方法在基本不影响预测框定位准确性的前提下,可有效提升预测框定位稳定性。     

基于搜索区域条件概率CNN的精确目标探测方法

针对传统目标探测方法多应用于低定位精度系统的情况,提出一种目标定位探测方法,以增强目标探测系统的定位精确度。确定候选框初始集;计算给定搜索区域的每行每列元素的条件概率,这些概率提供目标边界框位置的有用信息,根据概率情况,分别建立内外模型、边界模型和混合模型以返回感兴趣目标的边界框,实现定位;结合定位模型,利用卷积神经网络对目标进行训练探测。通过对PASCAL VOC和COCO数据集不同IoU阈值情况的实验,结果表明,与传统的方法相比,提出方法具有更高的探测准确率,可应用于高级目标探测系统。同时,利用滑动窗的方法确定候选框初始集,说明提出方法完全独立于传统的边界框回归方法。既简化了初始集的确定过程,同时保持较高的探测准确率。     

Hierarchical fusion and divergent activation based weakly supervised learning for object detection from remote sensing images

Object detection and location from remote sensing (RS) images is challenging, computationally expensive, and labor intense. Benefiting from research on convolutional neural networks (CNNs), the performance in this field has improved in the recent years. However, object detection methods based on CNNs require a large number of images with annotation information for training. For object location, these annotations must contain bounding boxes. Furthermore, objects in RS images are usually small and densely co-located, leading to a high cost of manual annotation. We tackle the problem of weakly supervised object detection under such conditions, aiming to learn detectors with only image-level annotations, i.e., without bounding box annotations. Based on the fact that the feature maps of a CNN are localizable, we hierarchically fuse the location information from the shallow feature map with the class activation map to obtain accurate object locations. In order to mitigate the loss of small or densely distributed objects, we introduce a divergent activation module and a similarity module into the network. The divergent activation module is used to improve the response strength of the low-response areas in the shallow feature map. Densely distributed objects in RS images, such as aircraft in an airport, often exhibit a certain similarity. The similarity module is used to improve the feature distribution of the shallow feature map and to suppress background noise. Comprehensive experiments on a public dataset and a self-assembled dataset (which we made publicly available) show the superior performance of our method compared to state-of-the-art object detectors.     

Adaptive video object proposals by a context-aware model

Most previous works focus on image object proposals while few on video object proposals. Besides, the existing explorations about video object proposals mainly concentrate on localizing the dominant object. In this paper, we aim at exploring a uniform framework for proposing multi-objects in videos no matter they are in the foreground or background. The method is derived from image object proposals, and makes best use of video characteristics. To achieve this task, we propose an adaptive context-aware model for video object proposals. First, spatial candidate windows are generated by the image method for acquiring the adequate bounding box samples. Temporal boxes are calculated by the motion based mapping. Considering the mapping loss, we define a box confidence coefficient contributing to keeping the proposal consistency and restraining the motion blur. The output proposal bounding boxes are ranked based on the scores calculated by the weighted scoring system. The proposed method is separately evaluated on the proposed multi-object dataset and the public dataset. The results compared with several state-of-the-arts show that our method has the most satisfactory overall performance for multi-object proposals in videos.     

Homogeneous bounding boxes as tools for intersection algorithms of rational bézier curves and surfaces

In the divide-and-conquer algorithm for detecting intersections of parametric rational Bézier curves (surfaces), we use bounding boxes in recursive rough checks. In this paper, we replace the conventional bounding box with a homogeneous bounding box, which is projectively defined. We propose a new rough check algorithm based on it. One characteristic of the homogeneous bounding box is that it contains a rational Bézier curve (surface) with weights of mixed signs. This replacement of the conventional bounding box by the homogeneous one does not increase the computation time.     

An Adaptive Spatial Subdivision of the Object Space for Fast Collision Detection of Animated Rigid Bodies

Collision detection tests between objects dominate run time simulation of rigid body animation. Traditionally, hierarchical bounding box tests are used to minimize collision detection time. But the bounding boxes do not take shapes of the objects into account which results in a large number of collision detection tests. We propose an adaptive spatial subdivision of the object space based on octree structure to rectify this problem. We also present a technique for efficiently updating this structure periodically during the simulation.     

Realtime Object-aware Monocular Depth Estimation in Onboard Systems

This paper proposes the object depth estimation in real-time, using only a monocular camera in an onboard computer with a low-cost GPU. Our algorithm estimates scene depth from a sparse feature-based visual odometry algorithm and detects/tracks objects’ bounding box by utilizing the existing object detection algorithm in parallel. Both algorithms share their results, i.e., feature, motion, and bounding boxes, to handle static and dynamic objects in the scene. We validate the scene depth accuracy of sparse features with KITTI and its ground-truth depth map made from LiDAR observations quantitatively, and the depth of detected object with the Hyundai driving datasets and satellite maps qualitatively. We compare the depth map of our algorithm with the result of (un-) supervised monocular depth estimation algorithms. The validation shows that our performance is comparable to that of monocular depth estimation algorithms which train depth indirectly (or directly) from stereo image pairs (or depth image), and better than that of algorithms trained with monocular images only, in terms of the error and the accuracy. Also, we confirm that our computational load is much lighter than the learning-based methods, while showing comparable performance.

     

YOLOv3与顶点偏移估计相结合的车牌定位

深层卷积神经网络(deep convolutional neural networks, DCNN)因其能够自动学习图像有效特征,被广泛应用于视觉目标检测.为克服DCNN目标检测算法大多因采用矩形检测框,而无法有效地应对非约束环境下倾斜性车牌的准确定位问题.提出一种可同时输出矩形目标检测框与关键点的车牌定位解决方案,并具体以YOLOv3所用网络为对象,通过扩展其输出维度,增设车牌顶点相对于矩形检测输出框角点的偏移量损失,在保留其高效计算性能的前提下,训练使其可同时输出矩形检测框及车牌顶点,实现精准定位.在广泛使用的大型非约束性车牌数据集CCPD上的实验结果显示,所提算法不仅可以准确检测车牌顶点,而且能够在Base,Tilt和Weather子集上取得99%以上的定位精度.该方法还可扩展至其他需同时输出目标检测框及关键点的应用领域,具有较好的应用价值.     

DSP: Discriminative Spatial Part modeling for Fine-Grained Visual Categorization

Different from the basic-level classification, the Fine-Grained Visual Categorization (FGVC) aims to classify objects belonging to the same species. Therefore, it is more challenging than the basic-level classification. Recently, significant advances have been achieved in FGVC. However, most of the existing methods require bounding boxes or part annotations for training and testing, resulting in limited usability and flexibility. To conquer these limitations, we aim to automatically detect the bounding boxes and parts for FGVC. The bounding boxes are acquired by transferring bounding boxes from training images to testing images. Based on the generated bounding boxes, we employ a multiple-layer Orientational Spatial Part (OSP) model to learn local parts for the object. To achieve more discriminative part modeling, the Discriminative Spatial Part (DSP) model is proposed to select the discriminative parts from OSP. Finally, we employ Convolutional Neural Network (CNN) as the feature extractor and train a linear SVM as the classifier. Extensive experiments on public benchmark datasets manifest the impressive performance of our method, i.e., classification accuracy achieves 79.8% on CUB-200-2011 and 85.7% on Aircraft, which are higher than many existing methods using manual annotations.     

基于包围盒和空间分解的碰撞检测算法

文中提出一种基于包围盒和空间分解的碰撞检测算法,用以解决软体的碰撞检测。算法使用AABB包围盒做初步检测,确定可能发生碰撞的物体。再根据包围盒的重叠情况缩小可能发生碰撞的区域,利用哈希表作为数据储存结构进行空间分解,将物体包围盒重叠区域的基本几何元素的空间网格映射到哈希表中,将碰撞区域缩小到基本几何元素,最后用基元碰撞检测找出具体碰撞点。由于前期AABB包围盒的处理减少了空间分解阶段需要映射的基本几何元素数量,该算法具有较高的运算速度。     

基于改进边界框回归损失的YOLOv3检测算法

YOLOv3检测算法中的边界框回归损失函数对边界框尺度敏感,且与算法检测效果评价标准交并比（IoU）之间的优化不具有强相关性,无法准确反映真值框与预测框之间的重叠情况,造成收敛效果不佳。针对上述问题,提出基于IoU的改进边界框回归损失算法BR-IoU。将IoU作为边界框回归损失函数的损失项,使不同尺度的边界框在回归过程中获得更均衡的损失优化权重。在此基础上,通过添加惩罚项最小化预测框与真值框中心点间围成的矩形面积,并提高预测框与真值框之间宽高比的一致性,从而优化边界框的回归收敛效果。在PASCAL VOC和COCO数据集上的实验结果表明,在不影响实时性的前提下,BR-IoU能够有效提高检测精度,采用BR-IoU的YOLOv3算法在PASCAL VOC 2007测试集上mAP较原YOLOv3算法和G-YOLO算法分别提高2.5和1.51个百分点,在COCO测试集上分别提高2.07和0.66个百分点。     

结合时空一致性的FairMOT跟踪算法优化

目的 视频多目标跟踪(multiple object tracking,MOT)是计算机视觉中的一项重要任务,现有研究分别针对目标检测和目标关联部分进行改进,均忽视了多目标跟踪中的不一致问题。不一致问题主要包括3方面,即目标检测框中心与身份特征中心不一致、帧间目标响应不一致以及训练测试过程中相似度度量方式不一致。为了解决上述不一致问题,本文提出一种基于时空一致性的多目标跟踪方法,以提升跟踪的准确度。方法 从空间、时间以及特征维度对上述不一致性进行修正。对于目标检测框中心与身份特征中心不一致,针对每个目标检测框中心到特征中心之间的空间差异,在偏移后的位置上提取目标的ReID(re-identification)特征;对帧间响应不一致,使用空间相关计算相邻帧之间的运动偏移信息,基于该偏移信息对前一帧的目标响应进行变换后得到帧间一致性响应信息,然后对目标响应进行增强;对训练和测试过程中的相似度度量不一致,提出特征正交损失函数,在训练时考虑目标两两之间的相似关系。结果 在3个数据集上与现有方法进行比较。在MOT17、MOT20和Hieve数据集中,MOTA(multiple object tracking accuracy)值分别为71.2%、60.2%和36.1%,相比改进前的FairMOT算法分别提高了1.6%、3.2%和1.1%。与大多数其他现有方法对比,本文方法的MT(mostly tracked)比例更高,ML(mostly lost)比例更低,跟踪的整体性能更好。同时,在MOT17数据集中进行对比实验验证融合算法的有效性,结果表明提出的方法显著改善了多目标跟踪中的不一致问题。结论 本文提出的一致性跟踪方法,使特征在时间、空间以及训练测试中达成了更好的一致性,使多目标跟踪结果更加准确。     

改进的基于AABB包围盒的碰撞检测算法

介绍了一种改进的基于AABB包围盒的碰撞检测算法,通过对对象不断的分割逐步构造出贴近对象的层次包围盒,在碰撞检测阶段对其逐层遍历以实现精确而快速的碰撞检测.实验结果表明,与层次包围球算法相比,该方法在构造二叉树和进行精确的碰撞检测时,性能都有较为明显的提高.     

Processing moving queries over moving objects using motion-adaptive indexes

This paper describes a motion-adaptive indexing scheme for efficient evaluation of moving continual queries (MCQs) over moving objects. It uses the concept of motion-sensitive bounding boxes (MSBs) to model moving objects and moving queries. These bounding boxes automatically adapt their sizes to the dynamic motion behaviors of individual objects. Instead of indexing frequently changing object positions, we index less frequently changing object and query MSBs, where updates to the bounding boxes are needed only when objects and queries move across the boundaries of their boxes. This helps decrease the number of updates to the indexes. More importantly, we use predictive query results to optimistically precalculate query results, decreasing the number of searches on the indexes. Motion-sensitive bounding boxes are used to incrementally update the predictive query results. Furthermore, we introduce the concepts of guaranteed safe radius and optimistic safe radius to extend our motion-adaptive indexing scheme to evaluating moving continual k-nearest neighbor (kNN) queries. Our experiments show that the proposed motion-adaptive indexing scheme is efficient for the evaluation of both moving continual range queries and moving continual kNN queries.     

旋转目标检测算法在卫星影像中的应用

近年来,深度学习在卫星影像目标检测领域得到了快速的发展,如何精准高效定位目标物体是卫星影像目标检测研究中的主要难点。提出了一种基于旋转矩形空间的YOLOv3改进算法来精准定位卫星影像目标,对原有网络进行改进,增加角度变换的数据预处理过程,防止实例角度变化对网络训练造成影响。使用双旋转坐标进行回归训练,增加了角度锚点,提高了网络对卫星目标的检测有效性。提出了基于旋转矩形空间的非极大值抑制改进算法,可以有效去除多余的旋转预测框。实验结果表明,改进YOLOv3算法相较于原始YOLOv3算法拥有更好的可视化效果,可以有效准确地定位卫星影像的目标物体,有效避免了密集场景下预测框的遮挡问题,在保证实时性的前提下,将均值平均精度提高了0.8个百分点。     

Action recognition using saliency learned from recorded human gaze

This paper addresses the problem of recognition and localization of actions in image sequences, by utilizing, in the training phase only, gaze tracking data of people watching videos depicting the actions in question. First, we learn discriminative action features at the areas of gaze fixation and train a Convolutional Network that predicts areas of fixation (i.e. salient regions) from raw image data. Second, we propose a Support Vector Machine-based recognition method for joint recognition and localization, in which the bounding box of the action in question is considered as a latent variable. In our formulation the optimization attempts to both minimize the classification cost and maximize the saliency within the bounding box. We show that the results obtained with the optimization where saliency within the bounding box is maximized outperform the results obtained when saliency within the bounding box is not maximized, i.e. when only classification cost is minimized. Furthermore, the results that we obtain outperform the state-of-the-art results on the UCF sports dataset.     

YOLO-wLU：考虑定位不确定性的目标检测算法

YOLOv3目标检测算法在检测目标时没有考虑边界框坐标定位存在的不确定性,因此有时不能得出正确的检测结果。针对此问题,提出YOLO-wLU（YOLO with Localization Uncertainty）算法。该算法借鉴深度学习中的不确定性思想,使用高斯分布函数建立边界框坐标的概率分布模型以考虑边界框坐标定位不确定性;设计新的边界框损失函数,在检测过程中移除定位不确定性较大的检测结果;通过融合周围边界框坐标信息提高了边界框坐标辨识结果的准确性。实验结果表明,该算法可有效减少误报率,提高检测精度;COCO数据集上测试结果显示,相比YOLOv3算法,该算法的mAP最高可提升4.1个百分点。