Jingling Wang

An Efficient Multi-object Tracking Method Using Multiple Particle Filters

Multiple objects tracking is an important and challenging issue, because of difficulties caused by variable number of objects and interaction of objects. In this paper, we present a distributed tracking approach based on Bayesian framework to avoid huge computational expenses involved in sampling from a joint state space. Single-object trackers easily suffer from false identities of objects after severe occlusions because of hidden first-order Markov hypotheses. To solve the problem, we define a transition matrix between consecutive frames to denote the occurrences and probabilities of dynamic events, such as continuation, appearance, disappearance, interaction and split associating current object detections and previous tracking results. Analyzing transition probabilities combined with position, direction and appearance, we can infer depth ordering of occlusions.The transition matrix is able to effectively guide multiple single-object particle filters to predict and update the state of objects. The simulations demonstrate that the proposed approach can initialize automatically and track varying number of objects with occlusions.

No-reference objective stereo video quality assessment based on visual attention and edge difference

No-reference objective stereo video quality assessment plays an important role in stereo video content production, especially to the sense of 2D-to-3D video conversion. The main problem needed to consider when designing an objective quality assessment model is to make the assessment result close to human perception. In this paper, a no-reference objective stereo video quality assessment method is proposed by taking consideration of visual attention and edge difference. For every frame in the given stereo video, the human interested regions and edge information are extracted separately. The edge information is used to calculate the blockiness, zero-crossing and disparity features for every block in the frames, and visual interested information determined the weights of each image block in the final assessment result calculation. Experimental results demonstrated the effectiveness of our proposed objective stereo video quality assessment method.

Human Action Recognition Using Multi-Velocity STIPs and Motion Energy Orientation Histogram

Local image features in space-time or spatio-temporal interest points provide compact and abstract representations of patterns in a video sequence. In this paper, we present a novel human action recognition method based on multi-velocity spatio-temporal interest points (MVSTIPs) and a novel local descriptor called motion energy (ME) orientation histogram (MEOH). The MVSTIP detection includes three steps: first, filtering video frames with multi-direction ME filters at different speeds to detect significant changes at the pixel level; thereafter, a surround suppression model is employed to rectify the ME deviation caused by the camera motion and complicated backgrounds (e.g., dynamic texture); finally, MVSTIPs are obtained with local maximum filters at multi-speeds. After detection, we develop MEOH descriptor to capture the motion features in local regions around interest points. The performance of the proposed method is evaluated on KTH, Weizmann, and UCF sports human action datasets. Results show that our method is robust to both simple and complex backgrounds and the method is superior to other methods that are based on local features.

Human action recognition using spatio-temoporal descriptor

A novel and efficient human action recognition method utilizing spatio-temporal interest point detector and 3D speed up robust features (3D SURF) descriptor is proposed. The spatio-temporal interest points are detected using two separate linear filters. Then 3D SURF descriptor is presented and demonstrated in detail to represent the local region around interest point. The experimental results on KTH and Weizmann dataset prove that the proposed method is superior to similar methods. Especially, compared with the popular 3DSIFT, 3D SURF is advantage in recognition rate and lower computation cost as well.

A Vehicle Detection Method in Night Outdoor Scenes Based on Spatio-temporal Bricks

A novel approach to detect vehicles in night outdoor scenes is proposed in this paper. By using online subspace learning, the preliminary foreground is obtained based on spatio-temporal bricks. Meanwhile, a new operator named 3DLBP is proposed to describe the local features of bricks. The final foreground is acquired by region-grow with 3DLBP and preliminary foreground. The results of experiment indicate that the proposed approach achieves higher performance in detecting vehicles in night outdoor scenes than common methods.

A Novel Particle Filtering Framework Using Genetic Monte Carlo Sampling

Particle degeneration is a key issue in the performance of a particle filter. In this paper we introduce genetic Monte Carlo into sampling process with the basic idea of solving particle degeneration by means of evolution thought. It is shown that the novel particle filtering framework can effectively eliminate particle degeneration and reduce its dependency on the particle validity. Furthermore, the new genetic particle filter can be optimized by three key genetic factors-selection, crossover and mutation probabilities. select individual, which is analogical with genetic operations (select, crossover and mutation). The essence of genetic operations shows that it is possible to use genetic operators to sample in the data space; this is also our motivation to bring out genetic Monte Carlo sampling. This paper introduces genetic Monte Carlo sampling method, and then uses it in the resampling step of particle filter with the basic idea of solving particle degeneration problem by means of evolution thought. Furthermore, in order to make the systematic performance to reach the maximum of accuracy and enhance the tracking robustness, we propose an approach for searching the best solution by adjusting the probabilities of select, crossover and mutation operation. Finally, we illustrate the operation of our framework by an example: One- dimensional growth model. Result demonstrates the effectiveness of our method comparing with the other particle filtering frameworks.

Target Tracking Based on Adaptive Particle Filter

This paper presents a method that can track non-rigid moving objects using adaptive particle filter based on spatiograms. Particle filters offer a probabilistic framework for dynamic state estimation and have proven to work well in target tracking. Two key components of particle filters are observation models and motion models. Firstly, because the observation model based on general color histograms discards the spatial information of images, the accuracy of the observation model is decreased. We adopt a proper observation model based on spatiograms which are histograms augmented with spatial means and covariances to capture a richer description of targets and increase robustness in tracking. Secondly, approximate fixed motion models used in practice, such as unrestricted random walking model with fixed noise variance, are not accurate enough. To overcome this problem, we adopt the adaptive multivariate autoregressive models which are computed via the regression analysis. The proposed adaptive motion models can adjust the model order, process noise variance and model parameters automatically. Also, the number of particles is adjusted automatically. The experiments show that the proposed algorithm can effectively track moving objects and increase the robustness in tracking. Its performance is compared with that of the general particle filtering algorithm to demonstrate the advantages of the new method.

Track Articulated Body Motion by Pose-Model Based Stochastic Search

Articulated body motion tracking remains a challenge problem as its high-dimension and the dependence between the joint freedoms. In this paper we present a novel framework to achieve articulated body motion (walking in this paper) tracking by the use of pose-model based stochastic search algorithm which partitions the state space into two pars(pose and noise) and then processes recursively. Besides, we also use marker-based Initialization to balance the limited information of the video, use path reform to improve the overall tracking performance. Experiment demonstrates the effectiveness of our method in solving the 2D articulated human movement tracking problem.

A Virtual Music Control System Based on Dynamic Hand Gesture Recognition

Gesture Recognition technology has been widely used in virtual reality and human-computer interaction. This paper proposed a virtual music control system based on dynamic hand gesture recognition. The system mode was mainly designed and realized by three modules including control terminal, client terminal and the server. By capturing the gesture image sequence via a cellphone camera, the system is able to recognize information characters of gestures such as number of fingers and movement of gesture trace. Control terminal generate different instructions and send them to client terminal via server. Relative experiments showed that the interaction system had good applicability and portability.

Bidirectional Markov Chain Monte Carlo Particle Filter for Articulated Human Motion Tracking

A novel framework of particle filter, named bidirectional Markov chain Monte Carlo particle filter (BMCMCPF), has been proposed to estimate articulated human movement state and action category jointly. Owing to the reason that we regard action category as the estimated state in our framework, firstly the motion models for every possible action are built via autoregressive modeling for the captured motion data with minimum distance. Meanwhile, the dynamic model and observation model also get coupled so that tracking and recognition can achieve synchronously. Then, the state estimation is completed by using the bidirectional Marko chain Monte Carlo sampling. BMCMCPF can not only improve the tracking performance as its global optimization property, but also smooth the joint’s movement trajectories to ensure the motion coordination. The experimental results on HumanEva datasets show that the effectiveness of BMCMCPF with unknown motion modality in solving the tracking problem.