K. Dick, François Charih, Yasmina Souley Dosso, L. Russell, J. Green
Google Street View and the emergence of self-driving vehicles afford an unprecedented capacity to observe our planet. Fused with dramatic advances in artificial intelligence, the capability to extract patterns and meaning from those data streams heralds an era of insights into the physical world. In order to draw appropriate inferences about and between environments, the systematic selection of these data is necessary to create representative and unbiased samples. To this end, we introduce the Systematic Street View Sampler (S3) framework, enabling researchers to produce their own user-defined datasets of Street View imagery. We describe the algorithm and express its asymptotic complexity in relation to a new limiting computational resource (Google API Call Count). Using the Amazon Mechanical Turk distributed annotation environment, we demonstrate the utility of S3 in generating high quality representative datasets useful for machine vision applications. The S3 algorithm is open-source and available at github.com/CU-BIC/S3 along with the high quality dataset representing power infrastructure in rural regions of southern Ontario, Canada.
{"title":"Systematic Street View Sampling: High Quality Annotation of Power Infrastructure in Rural Ontario","authors":"K. Dick, François Charih, Yasmina Souley Dosso, L. Russell, J. Green","doi":"10.1109/CRV.2018.00028","DOIUrl":"https://doi.org/10.1109/CRV.2018.00028","url":null,"abstract":"Google Street View and the emergence of self-driving vehicles afford an unprecedented capacity to observe our planet. Fused with dramatic advances in artificial intelligence, the capability to extract patterns and meaning from those data streams heralds an era of insights into the physical world. In order to draw appropriate inferences about and between environments, the systematic selection of these data is necessary to create representative and unbiased samples. To this end, we introduce the Systematic Street View Sampler (S3) framework, enabling researchers to produce their own user-defined datasets of Street View imagery. We describe the algorithm and express its asymptotic complexity in relation to a new limiting computational resource (Google API Call Count). Using the Amazon Mechanical Turk distributed annotation environment, we demonstrate the utility of S3 in generating high quality representative datasets useful for machine vision applications. The S3 algorithm is open-source and available at github.com/CU-BIC/S3 along with the high quality dataset representing power infrastructure in rural regions of southern Ontario, Canada.","PeriodicalId":281779,"journal":{"name":"2018 15th Conference on Computer and Robot Vision (CRV)","volume":"1 1","pages":"0"},"PeriodicalIF":0.0,"publicationDate":"2018-12-13","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"125521934","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":0,"RegionCategory":"","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}
Visual Teach and Repeat (VT&R) allows a robotic vehicle to navigate autonomously along a network of paths in the presence of illumination and scene changes. Traditionally, the system uses a stereo camera as the primary sensor for triangulating visual landmarks and often operates in highly textured outdoor environments. In this paper, we modify the VT&R system to use a monocular pipeline under the same framework, but also target indoor operation as a demonstration of a low-cost VT&R solution for warehouse logistics in a visually difficult environment. Unlike previous monocular VT&R solutions, we make no assumptions about the nature of the scene (e.g., local ground planarity). This allows the system to be readily deployable on more vehicles in a wider range of environments. To test the system, and motivated by a warehouse navigation application, an upward pointing camera is mounted on a Clearpath Husky ground vehicle. We demonstrate the vehicle is able to navigate with a 99.6% autonomy rate using such a system during 1.1 kilometers of driving, with an average scene depth that varies from 8-16 meters. The cross-track deviation from the taught path is less than 0.5 meters over 90% of the path, reaching a maximum of 0.85 meters and an average of 0.26 meters.
{"title":"Eye on the Sky: An Upward-Looking Monocular Teach-and-Repeat System for Indoor Environments","authors":"N. Zhang, M. Warren, T. Barfoot","doi":"10.1109/CRV.2018.00056","DOIUrl":"https://doi.org/10.1109/CRV.2018.00056","url":null,"abstract":"Visual Teach and Repeat (VT&R) allows a robotic vehicle to navigate autonomously along a network of paths in the presence of illumination and scene changes. Traditionally, the system uses a stereo camera as the primary sensor for triangulating visual landmarks and often operates in highly textured outdoor environments. In this paper, we modify the VT&R system to use a monocular pipeline under the same framework, but also target indoor operation as a demonstration of a low-cost VT&R solution for warehouse logistics in a visually difficult environment. Unlike previous monocular VT&R solutions, we make no assumptions about the nature of the scene (e.g., local ground planarity). This allows the system to be readily deployable on more vehicles in a wider range of environments. To test the system, and motivated by a warehouse navigation application, an upward pointing camera is mounted on a Clearpath Husky ground vehicle. We demonstrate the vehicle is able to navigate with a 99.6% autonomy rate using such a system during 1.1 kilometers of driving, with an average scene depth that varies from 8-16 meters. The cross-track deviation from the taught path is less than 0.5 meters over 90% of the path, reaching a maximum of 0.85 meters and an average of 0.26 meters.","PeriodicalId":281779,"journal":{"name":"2018 15th Conference on Computer and Robot Vision (CRV)","volume":"80 1","pages":"0"},"PeriodicalIF":0.0,"publicationDate":"2018-05-08","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"134140536","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":0,"RegionCategory":"","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}
Rezaul Karim, Md. Amirul Islam, N. Mohammed, Neil D. B. Bruce
In recent years, there have been significant advances in deep learning applied to problems in high-level vision tasks (e.g. image classification, object detection, semantic segmentation etc.) which has been met with a great deal of success. State-of-the-art methods that have shown impressive results on recognition tasks typically share a common structure involving stage-wise encoding of the image, followed by a generic classifier. However, these architectures have been shown to be vulnerable to the adversarial perturbations which may undermine the security of the systems supported by deep neural nets. In this work, initially we present rigorous evaluation of adversarial attacks on recent deep learning models for two different high-level tasks (image classification and semantic segmentation). Then we propose a model and dataset independent approach to generate adversarial perturbation and also the transferability of perturbation across different datasets and tasks. Moreover, we analyze the effect of different network architectures which will aid future efforts in understanding and defending against adversarial perturbations. We perform comprehensive experiments on several standard image classification and segmentation datasets to demonstrate the effectiveness of our proposed approach.
{"title":"On the Robustness of Deep Learning Models to Universal Adversarial Attack","authors":"Rezaul Karim, Md. Amirul Islam, N. Mohammed, Neil D. B. Bruce","doi":"10.1109/CRV.2018.00018","DOIUrl":"https://doi.org/10.1109/CRV.2018.00018","url":null,"abstract":"In recent years, there have been significant advances in deep learning applied to problems in high-level vision tasks (e.g. image classification, object detection, semantic segmentation etc.) which has been met with a great deal of success. State-of-the-art methods that have shown impressive results on recognition tasks typically share a common structure involving stage-wise encoding of the image, followed by a generic classifier. However, these architectures have been shown to be vulnerable to the adversarial perturbations which may undermine the security of the systems supported by deep neural nets. In this work, initially we present rigorous evaluation of adversarial attacks on recent deep learning models for two different high-level tasks (image classification and semantic segmentation). Then we propose a model and dataset independent approach to generate adversarial perturbation and also the transferability of perturbation across different datasets and tasks. Moreover, we analyze the effect of different network architectures which will aid future efforts in understanding and defending against adversarial perturbations. We perform comprehensive experiments on several standard image classification and segmentation datasets to demonstrate the effectiveness of our proposed approach.","PeriodicalId":281779,"journal":{"name":"2018 15th Conference on Computer and Robot Vision (CRV)","volume":"30 1","pages":"0"},"PeriodicalIF":0.0,"publicationDate":"2018-05-08","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"131898478","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":0,"RegionCategory":"","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}
Unmanned aerial vehicle (UAV) tracking is a challenging problem and a core component of UAV applications. CNNs have shown impressive performance in computer vision applications, such as object detection, image classification and so on. In this work, a locally connected layer is employed in a CNN architecture to extract robust features. We also utilize focal loss function to focus training on hard examples. Our CNN is first pre-trained offline to learn robust features. The training data is classified according to the texture, color, size of the target and the background information properties. In a subsequent online tracking phase, this CNN is fine-tuned to adapt to the appearance changes of the tracked target. We applied this approach to the problem of UAV tracking and performed extensive experimental results on large scale benchmark datasets. Results obtained show that the proposed method performs favorably against the state-of-the-art trackers in terms of accuracy, robustness and efficiency.
{"title":"Online Model Adaptation for UAV Tracking with Convolutional Neural Network","authors":"Zhuojin Sun, Yong Wang, R. Laganière","doi":"10.1109/CRV.2018.00053","DOIUrl":"https://doi.org/10.1109/CRV.2018.00053","url":null,"abstract":"Unmanned aerial vehicle (UAV) tracking is a challenging problem and a core component of UAV applications. CNNs have shown impressive performance in computer vision applications, such as object detection, image classification and so on. In this work, a locally connected layer is employed in a CNN architecture to extract robust features. We also utilize focal loss function to focus training on hard examples. Our CNN is first pre-trained offline to learn robust features. The training data is classified according to the texture, color, size of the target and the background information properties. In a subsequent online tracking phase, this CNN is fine-tuned to adapt to the appearance changes of the tracked target. We applied this approach to the problem of UAV tracking and performed extensive experimental results on large scale benchmark datasets. Results obtained show that the proposed method performs favorably against the state-of-the-art trackers in terms of accuracy, robustness and efficiency.","PeriodicalId":281779,"journal":{"name":"2018 15th Conference on Computer and Robot Vision (CRV)","volume":"8 1","pages":"0"},"PeriodicalIF":0.0,"publicationDate":"2018-05-08","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"133752651","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":0,"RegionCategory":"","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}
Alex D. Pon, Oles Andrienko, Ali Harakeh, Steven L. Waslander
Traffic light and sign detectors on autonomous cars are integral for road scene perception. The literature is abundant with deep learning networks that detect either lights or signs, not both, which makes them unsuitable for real-life deployment due to the limited graphics processing unit (GPU) memory and power available on embedded systems. The root cause of this issue is that no public dataset contains both traffic light and sign labels, which leads to difficulties in developing a joint detection framework. We present a deep hierarchical architecture in conjunction with a mini-batch proposal selection mechanism that allows a network to detect both traffic lights and signs from training on separate traffic light and sign datasets. Our method solves the overlapping issue where instances from one dataset are not labelled in the other dataset. We are the first to present a network that performs joint detection on traffic lights and signs. We measure our network on the Tsinghua-Tencent 100K benchmark for traffic sign detection and the Bosch Small Traffic Lights benchmark for traffic light detection and show it outperforms the existing Bosch Small Traffic light state-of-the-art method. We focus on autonomous car deployment and show our network is more suitable than others because of its low memory footprint and real-time image processing time. Qualitative results can be viewed at https://youtu.be/ YmogPzBXOw.
{"title":"A Hierarchical Deep Architecture and Mini-batch Selection Method for Joint Traffic Sign and Light Detection","authors":"Alex D. Pon, Oles Andrienko, Ali Harakeh, Steven L. Waslander","doi":"10.1109/CRV.2018.00024","DOIUrl":"https://doi.org/10.1109/CRV.2018.00024","url":null,"abstract":"Traffic light and sign detectors on autonomous cars are integral for road scene perception. The literature is abundant with deep learning networks that detect either lights or signs, not both, which makes them unsuitable for real-life deployment due to the limited graphics processing unit (GPU) memory and power available on embedded systems. The root cause of this issue is that no public dataset contains both traffic light and sign labels, which leads to difficulties in developing a joint detection framework. We present a deep hierarchical architecture in conjunction with a mini-batch proposal selection mechanism that allows a network to detect both traffic lights and signs from training on separate traffic light and sign datasets. Our method solves the overlapping issue where instances from one dataset are not labelled in the other dataset. We are the first to present a network that performs joint detection on traffic lights and signs. We measure our network on the Tsinghua-Tencent 100K benchmark for traffic sign detection and the Bosch Small Traffic Lights benchmark for traffic light detection and show it outperforms the existing Bosch Small Traffic light state-of-the-art method. We focus on autonomous car deployment and show our network is more suitable than others because of its low memory footprint and real-time image processing time. Qualitative results can be viewed at https://youtu.be/ YmogPzBXOw.","PeriodicalId":281779,"journal":{"name":"2018 15th Conference on Computer and Robot Vision (CRV)","volume":"101 1","pages":"0"},"PeriodicalIF":0.0,"publicationDate":"2018-05-08","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"124608282","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":0,"RegionCategory":"","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}
The counting of static targets on ground using an unmanned aerial vehicle (UAV) is proposed. To the best of our knowledge, this is the first paper to do such counting on-the-fly and autonomously. The flight path is programmed before take-off. The UAV captures images of the ground which are processed consecutively on-the-fly to count the number of targets along the flight path. Each image is processed using the proposed target-counting algorithm. First, targets' centers are detected in the current image, and second, the targets that were not covered in previous images are identified and counted. The performance of the algorithm depends on its ability to identify in the current image what targets were already counted in previous images, and this ability is affected by the limited accuracy of the UAV to stay on the flight path in the presence of wind. In the experimental evaluation, targets were distributed on ground on three different configurations: one line of targets along the flight path, parallel lines of targets at an angle with the flight path, and random. The accuracy of the target count was 96.0%, 88.9% and 91.9% respectively.
{"title":"Counting Static Targets Using an Unmanned Aerial Vehicle On-the-Fly and Autonomously","authors":"Dhruvil Darji, G. Vejarano","doi":"10.1109/CRV.2018.00037","DOIUrl":"https://doi.org/10.1109/CRV.2018.00037","url":null,"abstract":"The counting of static targets on ground using an unmanned aerial vehicle (UAV) is proposed. To the best of our knowledge, this is the first paper to do such counting on-the-fly and autonomously. The flight path is programmed before take-off. The UAV captures images of the ground which are processed consecutively on-the-fly to count the number of targets along the flight path. Each image is processed using the proposed target-counting algorithm. First, targets' centers are detected in the current image, and second, the targets that were not covered in previous images are identified and counted. The performance of the algorithm depends on its ability to identify in the current image what targets were already counted in previous images, and this ability is affected by the limited accuracy of the UAV to stay on the flight path in the presence of wind. In the experimental evaluation, targets were distributed on ground on three different configurations: one line of targets along the flight path, parallel lines of targets at an angle with the flight path, and random. The accuracy of the target count was 96.0%, 88.9% and 91.9% respectively.","PeriodicalId":281779,"journal":{"name":"2018 15th Conference on Computer and Robot Vision (CRV)","volume":"56 1","pages":"0"},"PeriodicalIF":0.0,"publicationDate":"2018-05-08","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"128315828","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":0,"RegionCategory":"","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}
With the increased use of cameras in robotic applications, this paper presents quantitative and qualitative assessment of the most prominent monocular tracking and mapping algorithms in literature with a particular focus on natural environments. This paper is unique in both context and methodology since it quantifies the performance of the state-of-the-art in Visual Simultaneous Localization and Mapping methods in the specific context where images mostly include vegetation. Finally, we elaborate on the limitations of these algorithms and the challenges that they did not address or consider when working in the natural environment.
{"title":"Survey of Monocular SLAM Algorithms in Natural Environments","authors":"Georges Chahine, C. Pradalier","doi":"10.1109/CRV.2018.00055","DOIUrl":"https://doi.org/10.1109/CRV.2018.00055","url":null,"abstract":"With the increased use of cameras in robotic applications, this paper presents quantitative and qualitative assessment of the most prominent monocular tracking and mapping algorithms in literature with a particular focus on natural environments. This paper is unique in both context and methodology since it quantifies the performance of the state-of-the-art in Visual Simultaneous Localization and Mapping methods in the specific context where images mostly include vegetation. Finally, we elaborate on the limitations of these algorithms and the challenges that they did not address or consider when working in the natural environment.","PeriodicalId":281779,"journal":{"name":"2018 15th Conference on Computer and Robot Vision (CRV)","volume":"156 1","pages":"0"},"PeriodicalIF":0.0,"publicationDate":"2018-05-08","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"123092573","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":0,"RegionCategory":"","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}
Nicholas Charron, Stephen Phillips, Steven L. Waslander
A common problem in autonomous driving is designing a system that can operate in adverse weather conditions. Falling rain and snow tends to corrupt sensor measurements, particularly for lidar sensors. Surprisingly, very little research has been published on methods to de-noise point clouds which are collected by lidar in rainy or snowy weather conditions. In this paper, we present a method for removing snow noise by processing point clouds using a 3D outlier detection algorithm. Our method, the dynamic radius outlier removal filter, accounts for the variation in point cloud density with increasing distance from the sensor, with the goal of removing the noise caused by snow while retaining detail in environmental features (which is necessary for autonomous localization and navigation). The proposed method outperforms other noise-removal methods, including methods which operate on depth image representations of the lidar scans. We show on point clouds obtained while driving in falling snow that we can simultaneously obtain > 90% precision and recall, indicating that the proposed method is effective at removing snow, without removing environmental features.
{"title":"De-noising of Lidar Point Clouds Corrupted by Snowfall","authors":"Nicholas Charron, Stephen Phillips, Steven L. Waslander","doi":"10.1109/CRV.2018.00043","DOIUrl":"https://doi.org/10.1109/CRV.2018.00043","url":null,"abstract":"A common problem in autonomous driving is designing a system that can operate in adverse weather conditions. Falling rain and snow tends to corrupt sensor measurements, particularly for lidar sensors. Surprisingly, very little research has been published on methods to de-noise point clouds which are collected by lidar in rainy or snowy weather conditions. In this paper, we present a method for removing snow noise by processing point clouds using a 3D outlier detection algorithm. Our method, the dynamic radius outlier removal filter, accounts for the variation in point cloud density with increasing distance from the sensor, with the goal of removing the noise caused by snow while retaining detail in environmental features (which is necessary for autonomous localization and navigation). The proposed method outperforms other noise-removal methods, including methods which operate on depth image representations of the lidar scans. We show on point clouds obtained while driving in falling snow that we can simultaneously obtain > 90% precision and recall, indicating that the proposed method is effective at removing snow, without removing environmental features.","PeriodicalId":281779,"journal":{"name":"2018 15th Conference on Computer and Robot Vision (CRV)","volume":"310 1","pages":"0"},"PeriodicalIF":0.0,"publicationDate":"2018-05-08","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"116758554","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":0,"RegionCategory":"","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}
In this work, a deflectometry based efficient semi-specular surface defect detection system is developed. This system, when integrated with a robotic arm, can detect defects on large and widely varying topological surfaces like a car bumper. A hybrid pipe line, that utilizes multi-threading, is designed to efficiently use resources and speedup the inspection process. Specific filters are also designed to eliminate spurious defects due to edges and acute curvature changes. The developed system was successful in consistently detecting various defects on large bumper parts with varying topology and color and can accommodate inherent ambient lighting and vibration issues.
{"title":"Automotive Semi-specular Surface Defect Detection System","authors":"Adarsh Tandiya, S. Akhtar, M. Moussa, Cole Tarry","doi":"10.1109/CRV.2018.00047","DOIUrl":"https://doi.org/10.1109/CRV.2018.00047","url":null,"abstract":"In this work, a deflectometry based efficient semi-specular surface defect detection system is developed. This system, when integrated with a robotic arm, can detect defects on large and widely varying topological surfaces like a car bumper. A hybrid pipe line, that utilizes multi-threading, is designed to efficiently use resources and speedup the inspection process. Specific filters are also designed to eliminate spurious defects due to edges and acute curvature changes. The developed system was successful in consistently detecting various defects on large bumper parts with varying topology and color and can accommodate inherent ambient lighting and vibration issues.","PeriodicalId":281779,"journal":{"name":"2018 15th Conference on Computer and Robot Vision (CRV)","volume":"34 1","pages":"0"},"PeriodicalIF":0.0,"publicationDate":"2018-05-08","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"120996873","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":0,"RegionCategory":"","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}
We present an approach to enhancing visual odometry and Simultaneous Localization and Mapping (SLAM) in the context of robot navigation by actively modulating the gaze direction to enhance the quality of the odometric estimates that are returned. We focus on two quality factors: i) stability of the visual features, and ii) consistency of the visual features with respect to robot motion and the associated correspondence between frames. We assume that local texture measures are associated with underlying scene content and thus with the quality of the visual features for the associated region of the scene. Based on this assumption, we train a machine-learning system to score different regions of an image based on their texture and then guide the robot's gaze toward high scoring image regions. Our work is targeted towards motion estimation and SLAM for small, lightweight, and autonomous air vehicles where computational resources are constrained in weight, size, and power. However, we believe that our work is also applicable to other types of robotic systems. Our experimental validation consists of simulations, constrained tests, and outdoor flight experiments on an unmanned aerial vehicle. We find that modulating gaze direction can improve localization accuracy by up to 62 percent.
{"title":"Gaze Selection for Enhanced Visual Odometry During Navigation","authors":"Travis Manderson, Andrew Holliday, G. Dudek","doi":"10.1109/CRV.2018.00025","DOIUrl":"https://doi.org/10.1109/CRV.2018.00025","url":null,"abstract":"We present an approach to enhancing visual odometry and Simultaneous Localization and Mapping (SLAM) in the context of robot navigation by actively modulating the gaze direction to enhance the quality of the odometric estimates that are returned. We focus on two quality factors: i) stability of the visual features, and ii) consistency of the visual features with respect to robot motion and the associated correspondence between frames. We assume that local texture measures are associated with underlying scene content and thus with the quality of the visual features for the associated region of the scene. Based on this assumption, we train a machine-learning system to score different regions of an image based on their texture and then guide the robot's gaze toward high scoring image regions. Our work is targeted towards motion estimation and SLAM for small, lightweight, and autonomous air vehicles where computational resources are constrained in weight, size, and power. However, we believe that our work is also applicable to other types of robotic systems. Our experimental validation consists of simulations, constrained tests, and outdoor flight experiments on an unmanned aerial vehicle. We find that modulating gaze direction can improve localization accuracy by up to 62 percent.","PeriodicalId":281779,"journal":{"name":"2018 15th Conference on Computer and Robot Vision (CRV)","volume":"47 1","pages":"0"},"PeriodicalIF":0.0,"publicationDate":"2018-05-08","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"122976278","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":0,"RegionCategory":"","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}
京公网安备 11010802042870号
京ICP备2023020795号-1
扫码关注我们
求助内容：
应助结果提醒方式：