Online First

Display Method:
Intelligent Vehicles Localization Based on Semantic Map Representation from 3D Point Clouds
ZHU Yuntao, LI Fei, HU Zhaozheng, WU Huawei
, Available online  
Abstract(6982) PDF(6051)
Abstract:
In order to improve the accuracy of node localization for intelligent vehicles,an intelligent vehicles localization method based on three-dimensional point clouds semantic map representation is proposed. The method is divided into three parts. Semantic segmentation based on 3D laser point clouds includes ground segmentation,traffic signs segmentation and pole-shaped target segmentation. Semantic map representation for intelligent vehicles uses segmented targets to project. Finally directional projections with weight,semantic roads and semantic codeing are generated. The codeing and global location from high-precision GPS make up representation model. Localization based on semantic representation model includes coarse localization from GPS matching and node localization from semantic coding matching. The experiments are carried out in three road scenes with different length and complexity,and the localization accuracy is 98.5%,97.6% and 97.8%,respectively. The results show that proposed method has high accuracy and strong robustness, which is suitable for different road scenes.
Companion Relationship Discovering Algorithm for Passengers in the Cruise Based on UWB Positioning
YAN Sixun, WU Bing, SHANG Lei, LYU Jieyin, WANG Yang
, Available online  
Abstract(6444) PDF(2589)
Abstract:
To accurately discover the companion relationship among passengers in the interior space of a cruise, UWB positioning is employed in the cruise to carry out on-board personnel location experiment. An improved Haussdorff-DBSCAN based scheme combined with indoor positional semantics is proposed to realize the trajectory clustering of the passenger trajectories, considering the characteristics of the UWB location data. Afterwards, the LSTM neural network is applied to predict the changing similarity of the suspected companions. Traditional Hausdorff algorithm does not consider the consistency of trajectory timing while calculating the trajectory similarity, and the introduction of positional semantic sequence can solve this problem well. In the first phase, the passenger trajectory data set is input to the improved Hausdorff-DBSCAN algorithm, and the clustering radius is determined according to the overall similarity threshold of trajectories. The outputs are the emerging clusters of passenger trajectories in the same companion group. In the second phase, the LSTM neural network takes the point similarity sequence with a fixed time window as the input to predict the point similarity value at the next time. The sequential change of passengers companion relationship is analyzed by the similarity threshold and prediction results. The validity of the presented algorithm is demonstrated by the trajectory data obtained from the passengers simulation on one deck of the cruise under study, which is modeled in Anylogic. The results indicate that the precision of the proposed algorithm reaches 0.92, the recall value reaches 0.95 and the F1 value is 0.934, which are at least 5.7 percent, 8.0 percent and 6.7 percent higher than the comparing algorithm. The LSTM neural network also shows a promising effect in predicting the changing trend of the similarity, for the loss is at a stable level of 3 to 4 percent.
Data Association Method Based on Descriptor Assisted Optical flow Tracking Matching
XIA Huajia, ZHANG Hongping, CHEN Dezhong, LI Tuan
, Available online  
Abstract(3433) PDF(923)
Abstract:
in the view of the problem that the positioning accuracy of visual inertial odometer using multi-state constrained Kalman filter(MSCKF) is easily affected by the abnormal value of feature point matching, a data association method based on descriptor assisted optical flow tracking matching is proposed. This method uses pyramid LK optical flow to track and match the feature points in the sequence image, then calculates the rbrief descriptor of each pair of matching points, judges the similarity of the descriptor according to the Hamming distance,and eliminates the abnormal matching points. In the experiment, the effectiveness of the proposed method is evaluated from two aspects:the subjective effect of feature point matching and positioning accuracy. The results show that the proposed method can effectively filter the abnormal values of image feature matching in dynamic scene. The image processed by this method is used for msckf motion solution,and the drift rate of position result is less than 0.38%, compared with the result of msckf algorithm without eliminating abnormal matching values,The improvement is 54.7%, and the single frame image processing time is about 39 ms.
Indoor Sign-based Visual Localization Method
HUANG Gang, CAI Hao, DENG Chao, HE Zhi, XU Ningbo
, Available online  
Abstract(7871) PDF(1193)
Abstract:
To solve the problem of localization calculation of intelligent vehicles and the mobile robot in the indoor traffic environment, by exploiting kinds of signs which existed in the indoor environment, a visual localization method is proposed through using BEBLID (Boosted Efficient Binary Local Image Descriptor) algorithm. The proposed method enforces the ability to characterize the whole image by improving the classic BEBLID. In this paper, the localization method consists of an offline stage and an online stage. In the offline stage, a scene sign map is created. In the online stage, the calculation progress is divided into 3 parts, which include holistic and local BEBLID method from current image and image in the scene sign map, closet sign site and closet image calculation by using KNN method, metric calculation by using coordinate information which is stored in the scene sign map. The experiment is conducted in three kinds of indoor scenes, including a teaching building, an office building, and an indoor parking lot. The experiment shows the scene sign recognition rate reached 90%, and the average localization error is less than 1 meter. Compared with the traditional method, the proposed method improves about 10% relative recognition rate with the same test set, which verified the effectiveness of the proposed method.
A Cooperative Map Matching Algorithm Applied in Intelligent and Connected Vehicle Positioning
CHEN Wei, DU Luyao, KONG Haiyang, FU Shuaizhi, ZHENG Hongjiang
, Available online  
Abstract(8038) PDF(951)
Abstract:
In order to achieve low-cost and high-precision vehicle positioning in the intelligent and connected environment,a cooperative map matching algorithm based on adaptive genetic Rao-Blackwellized particle filter is studied in this paper,improving the accuracy of vehicle positioning by using the real-time location data and road constraints of other connected vehicles. The adaptive genetic algorithm is introduced into the re-sampling process of the particle filter to increase the diversity of particles,so as to solve the problems of "particle degradation" and "particle exhaustion" that are prone to appear in traditional particle filter algorithms. Model of the algorithm is established and simulated. The positioning results under the traditional particle filter and Kalman smooth particle filter are compared,and the influence of the number of different connected vehicles on the positioning accuracy is analyzed. The experiment is completed in real-world and the performance of the algorithm is verified. The experimental results show that taking a typical intersection with four connected vehicles as an example,the range of position error of cooperative map matching is 1.67 m. It is only 41.03% and 56.80% of the traditional GNSS and the single map matching positioning results. At the same time,the circular error probable(CEP) of this algorithm is 1.06 m, which is 2.52 m higher than raw GNSS positioning result.