China 
Africa 
Explore chapters and articles related to this topic
Tracking a maneuvering target in clutter
Published in K. V. Ramachandra, Kalman Filtering Techniques for Radar Tracking, 2018
The probabilistic data association filter (PDAF) is a suboptimal bayesian algorithm which assumes that there is only one target of interest whose track has been initialized. In PDAF, the latest set of validated measurements are dealt with. It computes the probabilities of being correct for each validated measurement at the current time. It associates probabilistically all the neighbors to the target of interest. This probabilistic information used in PDAF accounts for the origin uncertainty. This is discussed in Section 10.4. PDAF is a recursive filter with fixed computational requirements and a minimum of modeling parameters. The only disadvantage of the original version of PDAF is that it cannot initiate or delete tracks. Recently, Colgrove, Davis, and Ayliffe [7] augmented the PDAF to include track initiation and deletion by adding in the association an event corresponding to “unobservable target” which can represent either a true target outside the sensor coverage or an erroneously hypothesized target which is equivalent to no target This technique enabled PDAF to initiate or delete tracks.
Application Tracking and Navigation
Published in Bin Jia, Ming Xin, Grid-based Nonlinear Estimation and Its Applications, 2019
Probabilistic Data Association Filter (PDAF) analyzes association probabilities of measurements and targets for each scan (measurement period). Specifically, PDAF calculates association probabilities for all measurements inside the validation gate. The procedure is given in Fig. 8.11 (Jwo et al. 2013).
Automatic traffic modelling for creating digital twins to facilitate autonomous vehicle development
Published in Connection Science, 2022
Shao-Hua Wang, Chia-Heng Tu, Jyh-Ching Juang
The detection of objects in point clouds is further decomposed into two procedures: detecting desired objects and calculating the moving tracks of the detected objects. The bounding box-based detecting algorithm is adopted to further detect the objects clustered in the previous step, where the L-shape bounding box algorithm (Zhang et al., 2017, June) is used and its result is appeared to be the shape of “L” from a top-down view. In addition to filtering out the unwanted objects (by using the pre-determined box length as the filtering threshold), the algorithm also helps in the estimation of the object orientation. The oriented boxes help further estimate the heading angles of the objects and are useful for tracking the paths of the objects.A tracker based on the algorithm of joint probabilistic data association filter (JPDAF) (Rachman, 2017) is used to track the movements of the multiple identified objects (boxes) in point clouds across a period of time. While it could involve the coalescing problem, where the closely spaced objects tend to coalesce over time during the analysis of the point clouds, it is not a critical issue in our study when trying to capturing the traffic flow of different types of moving objects. Later, some variant of the JPDAF algorithms, such as Set JPDAF, could be applied to avoid the track coalescence.
Deep learning-based pedestrian trajectory prediction and risk assessment at signalized intersections using trajectory data captured through roadside LiDAR
Published in Journal of Intelligent Transportation Systems, 2023
Shanglian Zhou, Hao Xu, Guohui Zhang, Tianwei Ma, Yin Yang
Deep Learning (DL)-based methodologies have rapidly emerged for pedestrian trajectory prediction in recent years. Compared to other types of methods, DL-based methodologies can directly learn from data and make self-adaptations, owing to their ability to achieve a high level of feature abstraction through a hierarchical layout, thus mitigating the uncertainties from human intervention or prior user input. Methodologies using the vanilla Recurrent Neural Network (RNN) and its variants such as Long Short-Term Memory (LSTM) network and Gated Recurrent Unit (GRU) for pedestrian trajectory prediction have been reported in the literature (Ahmed et al., 2019; Li et al., 2019; Ridel et al., 2018). Sun et al. (2018) developed a DL-based approach for pedestrian trajectory prediction from LiDAR data collected by mobile service robots, in which they trained a shared-triple-layer LSTM network to learn from long-term temporal information and short-term pedestrian pose observations. Xue et al. (2019) proposed a methodology based on an encoder-decoder LSTM network to predict pedestrian trajectories using data captured by a vehicle-mounted LiDAR sensor. The encoder with a two-stream layout was designed to extract features from vehicle and pedestrian trajectories, which were then fused through the decoder for pedestrian trajectory prediction. Besides, Zhang et al. (2021) proposed a deep architecture based on Convolutional Neural Networks (CNNs) to predict pedestrian trajectories using LiDAR data captured in autonomous driving scenes. Zhang et al. (2022) proposed a joint detection and tracking scheme to predict the trajectories of moving objects, which involves two parallel procedures: the PointVoxel-RCNN network was employed to detect vehicle and pedestrian objects, while the Unscented Kalman Filter (UKF) and Joint Probabilistic Data Association Filter (JPDAF) were utilized for trajectory prediction. Wang et al. (2022) developed a deep learning-based multi-sensor fusion approach to detect and track traffic objects. In Wang et al. (2022), two-dimensional (2D) trajectories were predicted by the improved YOLOv5 network from image data and then fused with 3D trajectories from roadside LiDAR data by applying the improved PointRCNN network.