专业英语之文献语言点评论.doc

科技英语阅读与写作题 目 Pedestrian Detection and Tracking Using Deformable Part Models and Kalman Filtering 学 院 专 业 姓 名 soppa=face like this 学 号 parterner姓 名 学 号 Pedestrian Detection and Tracking Using Deformable Part Models and Kalman FilteringXue Fan, Shubham Mittal, Twisha Prasad, Suraj Saurabh and Hyunchul ShinReceived: February 23, 2013 /Accepted: March 22, 2013 /Published：July 31, 2013Abstract: Pedestrian detection techniques are important and challenging especially for complex real world scenes. They can be used for ensuring pedestrian safety, ADASs(advance driver assistance systems) and safety surveillance systems. In this paper, we propose a novel approach for multi-person tracking-by-detection using deformable part models in Kalman filtering framework. The Kalman filter is used to keep track of each person and a unique label is assigned to each tracked individua1. Based on this approach, people can enter and leave the scene randomly. We test and demonstrate our results on Caltech Pedestrian benchmark. which is two orders of magnitude larger than any other existing datasets and consists of two orders of magnitude，表示“两个数量级”，科技文中常用的数量级比较，“larger than”也表明比较级，“any other”常用于比较级中译为“其他任何”。pedestrians varying widely in appearance, pose and scale. Complex situations such as people occluded by each other are handled gracefully and individual persons can be tracked correctly after a group of people split. Experiments confirm the real-time performance and robustness of our system, working in complex scenes. Our tracking model gives a tracking accuracy of 72.8% and a tracking precision of 82.3%. We can further reduce false positives by 2.8%, using Kalman filtering.Key words：multi-person tracking, deformable part models, data association, Kalman Filtering, pedestrian detection.1. IntroductionPedestrian detection and tracking are important in a wide range of applications in our daily lives, such as“such as 比如、诸如” 相当于like或forexample。 surveillance systems, airport security, automatic driving, and driver assistance systems in high-end cars, human-robot interaction, and英语中多个短语并列时，在最后一个处用and 连接，其余的用逗号连接，注意英语中无顿号的用法，这点注意与汉语的差别。 assistance for senior citizens. Detecting pedestrians in images is a challenging task owing to various styles of clothing in appearance and huge possible postures. Significant research has been devoted to detecting, locating, and tracking people in images and videos. The goal of this paper is to enable reliable multi-person tracking from a moving platform in real-world scenarios. In this paper, the authors present an integrated tracking-by-detection framework which is able to detect and track multiple persons even in challenging scenarios such as instances where the objects are partially occluded for extended periods of time. We integrated formable part based models with Kalman filtering. We use the detector of Ref. 1 and modify the source code to train it on Caltech Pedestrian Dataset2The main challenge when using the pedestrian detector for tracking is that the detector output is unreliable. The detection output sometimes consists of some false positives and missed detections. Thus, the resulting association problem between detections and targets is difficultSeveral recent algorithms address this problem by optimizing detection assignments over a large temporal window in an offline step3,4With the improvement of detection algorithms 1，both in accuracy and computational feasibility，tracking-by-detection is one of the most popular concepts for tracking5Due to the high amount of false positives and missing detections in the output of the detector，it is necessary to incorporate temporal contextRecent tracking approaches6 try to associate the detections and track objects from uncalibrated single cameraRecently，pedestrian detection methods have achieved great improvements7,8， which have allowed the scientific community to focus on associating tracklets whichWhich引导主语从句 are initially linked by detection responses in consecutive frames Next，short tracklets are associated to form longer ones by maximizing the probabilities globally using the similarity and discrimination scores both from appearance and motion modelsBy using the information of the previous，current，and future frames，these methods rely on the spatio-temporal restriction to improve the results with different types of features fromresult from 也是常用短语，译为“起因于；由造成” simple object signatures such as height，width to more complicated ones like local gradient intensity featuresThe most important factor to improve the tracking results is building the affinity scores between two tracklets to associate tracklets to the right target. Our paper makes the following contributions：(1) We test our model on the Caltech Pedestrian Dataset (the biggest pedestrian dataset available) and demonstrate impressive results even in heavy occluded scenes；(2)we show that our method achieves high tracking accuracy by successfully associating data within consecutive frames to overcome the problem of unreliable detectionThe rest of the paper is organized as followsIn Section2，a brief review of the related work is presentedDeformable part models are shown in Section3Kalman filtering is presented in Section4Data Association is presented in Section5Results are shown in Section6Finally，the conclusions and possible future works are summarized in Section72Related Works2.1 Tracking-by-DetectionMany tracking methods rely on background subtraction from one or several static cameras9,10But，recent improvements in object detection have stimulated the interest in combining tracking and detectionThe combination of object detectors and Kalman filtering results in algorithms that are more suitable for time-critical，online applicationsOn the other hand，state-of-the-art object detectors all build up some form of confidence density as one stage of their pipeline, which can be used instead as a graded observation model to handle difficult situations more 比较级，多音节比较级用more，robustly译为“要用体力地、强壮地”。robustlyPrevious algorithms that exploit this intermediate output have been developed primarily for single-target tracking (mostly of faces) and have not been evaluated thoroughly for multiple interacting targetsRelying on the detector confidence in every situation can cause tracking errors，particularly during occlusions between interacting targets and in complex cluttered scenesThis work presents a method to use this unreliable information source for robust multi-person tracking2.2 Data AssociationUsing independent trackers require solving a data association problem to assign detections to targetsA large number of strategies are available to solve the data association problemClassical approaches include the JPDAF(joint probabilistic data association filter)11 and MHT (multi hypotheses tracking ) 12MHT attempts to keep track of all the possible association hypotheses over timeThis is an NP-hard problem, since the number of association hypotheses grows exponentially over timeThus methods are required to reduce the computational complexityJPDAFs instead try to make the best possible assignment in each time step by jointly considering all possible associations between targets and detections，to the cost of an exponentially increasing complexity3Deformable Part ModelsIn this paper we use a pedestrian detection system based on mixtures of multi-scale deformable part models1The structure of deformable pedestrian detector model and a detection example are shown in Fig.1. These models are trained using a discriminative procedure that only require bounding boxes for the object in a set of images. The system lies heavily on new methods for training of classifiers that make use of latent information. It also relies heavily on efficient methods for matching deformable models to images.The core ideas of deformable part-based models boil down to there factors: (1) A deformable part representation for pedestrian; (2) An efficient matching process; (3) Latent SVM, a discriminative training method. A star-structured part-based model is defined which is composed of a root filter, n (usually six) part filters, and associated deformation parameters. An efficient matching process based on dynamic programming and generalized distance transforms is programming and generalized distance transforms is proposed. To detect objects in an image we compute an overall score for each root location according to the best possible placement of the parts. Finally, a latent SVM training process is formulated to train a mixture of star models from bounding box ground truth.Part-based models focus on parts of an object instead of trying to capture a global pattern of an object with one template. Thereby, they can provide more flexible and robust representations. The resulting system is both efficient and accurate, achieving state-of-the-art INRIA Person dataset.4. Kalman FilteringMathematically, Kalman filter is an estimator that predicts and corrects the optimal state is found with smallest possible variance error, recursively.In Kalman filter, we consider a behavior of the object, where subscript k indicates the discrete time. The objective is to compute posteriori state estimate from the measurement and apriori state estimate . The measurement and aprior state estimate . The overview of our framework is shown in Fig.2. In the following, we give a mathematical description of Kalman filter. 4.1 Process EquationThe Kalman filter address the general problem of trying to estimate the state of a discrete-time controlled process that is governed by the linear stochastic difference equation:Where A represents the state transition model and is the state at time k. the random variable is the Gassian process noise N() with following normal probability distribution p(w)：4.2 Measurement Equationwhere，H represents the observation model and zk is the measurement observed at time kThe random variable represents the Gaussian measurement noise N() with following normal probability distribution p(v)：4.3 Time Update EquationsWe defineto be our a priori state estimate at step k given knowledge of the process prior to step k, and to be priori estimate error covariance. and are obtained for the next time step in following equations：These equations show how to project the state and covariance estimates forward from time step k-1 to step k4.4 Measurement Update EquationsWe defineto be our aposteriori state estimate at step k given measurement and apriori estimate ，and to be posteriori estimate error covarianceThe objective is to find an equation that compute a posteriori estimating which is a linear combination of the apriori estimate and the new measurement These equations are given below：Whereis the Kalman gain一、语言点 demonstrate on 译为“在上演示”，做谓语；which引导的非限定性定语从句，修饰本实验所依靠的基准的特性；consists of 译为“由组成”，varying 现在分词做状语，修饰pedestrians，说明行人在外观、姿势和数量上的变化。 confirm 译为“证实”，科技文中常用表述“实验证实了.”用“Experiments confirm”表达；the real-time performance 译为“实时性能”，科技文中常用到的实时性用“real-time”表示；working分词短语做伴随状语；in complex scenes译为“在复杂场景中”。 in a wide range of applications“在中有广泛应用”，可用于文章开头对某物进行总体介绍；in our daily lives译为“在我们的日常生活中”，这两句常用于总体介绍某事物的用途。特别要注意在不同领域的专业说法：surveillance systems：服务系统 airport security：机场安全 automatic driving：自动驾驶 driver assistance systems：驾驶员辅助系统 high-end cars：高端汽车 human-robot interaction：人机交互 assistance for senior注意senior除了有“大四”的含义外，还有“高级的；年长的；资深的”含义。 citizens：帮助老年人辅助老年人。 Detecting pedestrians动名词短语做主语；challenging task 译为“有挑战的任务”，常用于科技文中对某事物的研究难度；owing to表示原因，译为“由于、因为”； various styles of译为“风格迥异的”，修饰后面的着装的姿势。 when引导时间状语从句，using the pedestrian detector动名短语词做主语，for tracking 表目的，英文中常用不定式to do sth 或者for doing sth来表示做某事的目的；thatThe main challenge做形式主语，真正的主语是is that后面引导的部分，次句式常用于英文中，为了避免头重脚轻。后面的句子做表语。 With the improvement of，译为“随着的改进”，做状语；bothand常用于连接两部分，“都”；tracking-by-detection 动名词短语做主语，by doing sth 译为“通过(方式)”； one of the most是最高级的用法，译为“最之一”。 Due to 译为“由于，归咎于”，表原因，it做形式主语，避免头重脚轻，it is necessary to do sth，“做.很有必要”，此句式常用于英文中陈述某事。 Recently表明此句时态应用现在完成时，即have achieved；achieve great improvements译为“取得了极大的进步”，英文中取得什么样的成绩、达到什么效果等常用achive表示，尽量少用get；第一个Which引导非限制性定语从句，focus on译为“集中，焦点关注于.”。 by + -ing 表示方式，译为“通过手段”；rely on译为“依靠、依赖”；restriction to注意此处的动词原型是res