Prior Motion Model Research Articles

Motion priors based on goals hierarchies in pedestrian tracking applications

In this paper, the problem of automated scene understanding by tracking and predicting paths for multiple humans is tackled, with a new methodology using data from a single, fixed camera monitoring the environment. Our main idea is to build goal-oriented prior motion models that could drive both the tracking and path prediction algorithms, based on a coarse-to-fine modeling of the target goal. To implement this idea, we use a dataset of training video sequences with associated ground-truth trajectories and from which we extract hierarchically a set of key locations. These key locations may correspond to exit/entrance zones in the observed scene, or to crossroads where trajectories have often abrupt changes of direction. A simple heuristic allows us to make piecewise associations of the ground-truth trajectories to the key locations, and we use these data to learn one statistical motion model per key location, based on the variations of the trajectories in the training data and on a regularizing prior over the models spatial variations. We illustrate how to use these motion priors within an interacting multiple model scheme for target tracking and path prediction, and we finally evaluate this methodology with experiments on common datasets for tracking algorithms comparison.

Particle-filter-based Pose Estimation from Controlled Motion with Application to Visual Servoing

In this paper, we present a Bayesian algorithm based on particle filters to estimate the camera pose for vision-based control. The state model is represented as a relative camera pose between the current and initial camera frames. The particles in the prior motion model are drawn using the velocity control signal collected from the visual controller of the robot. The pose samples are evaluated using an epipolar geometry measurement model and a suitable weight is associated with each sample. The algorithm takes advantage of the a priori knowledge about motion, i.e., the velocity computed by the visual servo control, to estimate the magnitude of the translation in addition to its direction, hence producing a full camera motion estimate. Its application to position-based visual servoing is demonstrated. Experiments are carried out using a real robot setup. The results show the efficiency of the proposed filter over the motion measurements of the robot. In addition, the filter was able to recover the split performed by the robot joints.

Probe localization for freehand 3D ultrasound by tracking skin features.

Ultrasound probe localization with respect to the patient's body is essential for freehand three-dimensional ultrasound and image-guided intervention. However, current methods for probe localization generally involve bulky and expensive equipment. In this paper, a highly cost-effective and miniature-mobile system is described for 6-DoF probe localization that is robust to rigid patient motion. In this system, skin features in the scan region are recorded at each ultrasound scan acquisition by a lightweight camera rigidly mounted to the probe. A skin map is built based on the skin features and optimal probe poses are estimated in a Bayesian probabilistic framework that incorporates a prior motion model, camera frames, and ultrasound scans. Through freehand scanning on three different body parts, it is shown that on average, for every probe travel distance of 10 mm, the translational and rotational errors are 0.91 ± 0.49 mm and 0.55 degrees ± 0.17 degrees, respectively. The 3D reconstructions were also validated by comparison with real ultrasound scans.

Data modeling and query processing for distributed surveillance systems

This paper presents data modeling and query processing for distributed surveillance systems. We define a metadata rule to search and manage information for distributed or heterogeneous surveillance systems. For human activity recognition, we propose a method that classifies these actions separately from complicated activities as a sequence of basic activities. In addition, we define the domain and range of relations based on the relationship between elements. Furthermore, we describe the state descriptors to represent an image sequence. To address the interaction of multiple objects, we classify human actions into symmetric or asymmetric actions. The prior motion model and the inference approach are applied adaptively according to environments. We define the grammar for the representation of the surveillance video and specify different query criteria for surveillance video retrieval. In the experiments, we show the prototype system that provides event detection, object identification, object tracking, face recognition, and activity recognition.

TH-E-218-05: Prediction of Respiratory Motion from Single Daily 3D Image Using Prior Model of Motion and Anatomic Variations

Purpose: Respiratory motion may vary significantly from the planning stage to the time of treatment, leading to challenges in motion‐based treatment intervention. The purpose of this study was to develop a computational framework allowing for accurate prediction of daily respiratory motion from single 3D daily image under drastic inter‐fractional respiratory variations and anatomic changes, by taking advantage of prior knowledge of motion and anatomic relationship. Methods: Deformable image registration (DIR) was first performed across planning 4D‐CT to obtain a set of displacement vector field (DVF), which were further modeled by Principal Component analysis (PCA) to learn a prior motion model, namely a subspace spanned by principal bases. Subsequently, DIR was employed again across vowels of planning and daily images to calculate the DVF for inter‐fractional motion, which was used to spatially map the learned prior motion subspace (principal bases) onto the grid of daily image. Finally, we estimated the component of inter‐fractional respiratory variations by linear projection of inter‐fractional DVF onto mapped motion subspace. The resultant projections were utilized to compensate the reconstruction of daily respiratory motion. The reconstructed respiratory motion will be useful for various image‐guided treatment intervention strategies. Results: We applied the proposed framework for the prediction of respiratory motion from single daily CBCT and/or free‐breathing conventional CT using patients from a proton protocol undergoing weekly 4D‐CT evaluation. The accuracy of our model was visually and quantitatively confirmed by the excellent agreement between predicted contours by our model and physician approved ones. For a case with large tumor shrinkage, a DICE score of 90.2% has been achieved; even for another case with drastic respiratory motion difference (around 2cm) from planning stage, the agreement is also close to 90%. Conclusions: We have proposed and validated a novel framework to predict respiratory motion from single 3D daily image.

SP-0029 EXPLOITING HYPOXIA DEPENDENT DNA REPAIR DEFECTS

model for MC CBCT reconstruction. The accuracy of the MC CBCT based correction strategy was first quantified using a 4D phantom. For each fraction, a first MC CBCT is acquired to align the target to the planned position, a second scan to verify this correction and two additional scans are acquired concurrent with VMAT delivery to monitor intra-fraction motion. The first scan of each fraction was reconstructed as a 4D MC CBCT to assess differences between the prior motion model and actual patient motion. Subsequent scans were 3D MC CBCT if the residual target motion amplitude ≤ 8 mm and otherwise regular 4D CBCT were used. Results: Phantom experiments demonstrated alignment accuracy better then 1 mm/1° compared to 4D CBCT and a residual amplitude on 4D MC CBCT ≤ 2 mm. Six patients have been evaluated so far using MC CBCT protocol since it was clinically implemented in October 2011 on one of our treatment machines. The mean (SD) tumour amplitude derived from the 4D planning CT was 2 (2) mm (LR), 5 (4) mm (CC) and 5 (3) mm (AP). A total of 25 fractions were analyzed. In one patient, the residual tumour amplitude exceeded 8mm; so we reverted to a 4DCBCT guidance protocol. The residual tumour amplitude derived from the first scan of each fraction using 4D MCCBCT scans for the remaining patients was 1.4 (0.6) mm (LR), 1.6 (0.7) mm (CC) and 1.7 (0.5) mm (AP), showing validaty of the motion compensation (figure 1).

Bayesian Inference of Models and Hyperparameters for Robust Optical-Flow Estimation

Selecting optimal models and hyperparameters is crucial for accurate optical-flow estimation. This paper provides a solution to the problem in a generic Bayesian framework. The method is based on a conditional model linking the image intensity function, the unknown velocity field, hyperparameters, and the prior and likelihood motion models. Inference is performed on each of the three levels of this so-defined hierarchical model by maximization of marginalized a posteriori probability distribution functions. In particular, the first level is used to achieve motion estimation in a classical a posteriori scheme. By marginalizing out the motion variable, the second level enables to infer regularization coefficients and hyperparameters of non-Gaussian M-estimators commonly used in robust statistics. The last level of the hierarchy is used for selection of the likelihood and prior motion models conditioned to the image data. The method is evaluated on image sequences of fluid flows and from the "Middlebury" database. Experiments prove that applying the proposed inference strategy yields better results than manually tuning smoothing parameters or discontinuity preserving cost functions of the state-of-the-art methods.

TU-C-214-05: Digital Synthesizing High-Quality Daily 4D-CBCT from Conventional 3D-CBCT Using Prior Knowledge of Motion and Anatomy Relationships

Purpose: Conventional in-room Cone-beam CT (CBCT) lacks explicit motion representation and usually has poor image quality and inaccurate CT numbers for target delineation and/or adaptive planning. While in-room 4D- CBCT image acquisition is still time-consuming and suffers same poor image quality issue, the goal of this study was to develop a computational procedure to digitally synthesize high-quality daily 4D-CBCT images by taking advantage of the prior knowledge of motion and appearance defined in the planning 4D-CT dataset.Methods: We first estimated the respiratory motion from planning 4D-CT by performing deformable image registration between the averaged planning 4D-CT to each of the phases. The resultant displacement motion vectors were further analyzed by Principal Component analysis (PCA) to construct a patient-specific low-dimensional motion model. Subsequently, dense correspondences were established across voxels of daily 3D-CBCT and averaged 4D-CT. Given voxel-by-voxel correspondences, both the learned prior motion model and CT numbers from the planning CT were spatially mapped onto the grid of daily CBCT. The transferred motion model allows us to synthesize 4D-CBCT images with a substantially improved image quality compared to the original 3D-CBCT. The synthesized 4D images possess explicit patient motion while maintains the geometric accuracy of patientˈs anatomy at the time of treatment. Results: We validated our model by comparing the synthesized 4D-CBCT against the 4D-CT dataset acquired in the same day from a protocol patient undergoing daily in-room CBCT setup and weekly 4D-CT for treatment evaluation. The correctness of contours and checkboard image fusion confirmed the accuracy of our model. In another patient example with large tumor shrinkage, we have observed excellent agreement of contours in different motion phases between the synthesized and acquired scans. Various imaging artifacts were greatly suppressed and soft-tissue visibility was drastically enhanced. Conclusions: We have proposed and validated a novel framework to render high-quality daily 4D-CBCT.

A neural architecture for visual motion perception: Group and element apparent motion

A neural network model of motion segmentation by visual cortex is described. The model's properties are illustrated by simulating on the computer data concerning group and element apparent motion, including the tendency for group motion to occur at longer ISIs and under conditions of short visual persistence. These phenomena challenge recent vision models because the switch between group and element motion is determined by changing the timing of image displays whose elements flash on and off but do not otherwise move through time. The model clarifies the dependence of short-range and long-range motion on a spatial scale. Its design specifies how sustained response cells and transient response cells cooperate and compete in successive processing stages to generate motion signals that are sensitive to direction-of-motion, yet insensitive to direction-of-contrast. Properties of beta motion, phi motion, gamma motion, and Ternus motion are explained. A number of prior motion models are clarified, transformed, and unified, including the Reichardt model, Marr-Ullman model, Burt-Sperling model, Nakayama-Loomis model, and NADEL model. Apparent motion and real motion generate testably different model properties. The model clarifies how preprocessing of motion signals by a motion OC Filter is joined to long-range cooperative motion mechanisms in a motion CC Loop to control phenomena such as induced motion, motion capture, and motion after effects. The total model system is a motion Boundary Contour System (BCS) that is computed in parallel with the static BCS of Grossberg and Mingolla before both systems cooperate to generate a boundary representation for 3-D visual form perception.

Opponent mechanisms in vision : Deborah Walters. Department of Computer Science, State University of New York at Buffalo, Buffalo, NY 14260 USA

A neural network model of motion segmentation by visual cortex is described. The model's properties are illustrated by simulating on the computer data concerning group and element apparent motion, including the tendency for group motion to occur at longer ISIs and under conditions of short visual persistence. These phenomena challenge recent vision models because the switch between group and element motion is determined by changing the timing of image displays whose elements flash on and off but do not otherwise move through time. The model clarifies the dependence of short-range and long-range motion on a spatial scale. Its design specifies how sustained response cells and transient response cells cooperate and compete in successive processing stages to generate motion signals that are sensitive to direction-of-motion, yet insensitive to direction-of-contrast. Properties of beta motion, phi motion, gamma motion, and Ternus motion are explained. A number of prior motion models are clarified, transformed, and unified, including the Reichardt model, Marr-Ullman model, Burt-Sperling model, Nakayama-Loomis model, and NADEL model. Apparent motion and real motion generate testably different model properties. The model clarifies how preprocessing of motion signals by a motion OC Filter is joined to long-range cooperative motion mechanisms in a motion CC Loop to control phenomena such as induced motion, motion capture, and motion after effects. The total model system is a motion Boundary Contour System (BCS) that is computed in parallel with the static BCS of Grossberg and Mingolla before both systems cooperate to generate a boundary representation for 3-D visual form perception.