Shiqiang Hu

Detection-guided multi-target Bayesian filter

Multi-target Bayesian filter in the framework of finite set statistics (FISST) and its approximations, including probability hypothesis density (PHD) filter and cardinalized probability hypothesis density (CPHD) filter, are elegant methods for multi-target tracking by jointly estimating the number of targets and their states from a sequence of noisy and cluttered observation sets. PHD filter and CPHD filter can deal with the tracking scenario involving the surviving targets, the spawned targets, and the spontaneous births. One of the limitations of PHD and CPHD filter is that it is assumed that intensities of spontaneous birth targets are known at the initialization stage. To address the problem, a track initiation technique is proposed to detect the position unknown birth targets and is hybridized with PHD and CPHD filter. Once new targets are detected, the position estimates are employed to form intensities of spontaneous births for starting PHD and CPHD filter. Simulation results demonstrate that the proposed tracker can adaptively and efficiently track multiple targets especially in scenarios with birth targets of unknown position, which the PHD and CPHD filter are unable to do on their own.

Visual Tracking via Constrained Incremental Non-negative Matrix Factorization

This letter presents a novel visual tracking algorithm by using Incremental Non-negative Matrix Factorization (INMF) and dual ℓ1-norm constraints. Firstly, we introduce one ℓ1 regularization into the NMF reconstruction, which enables appearance model to tolerate different noises to some extent. Meanwhile, we enforce another ℓ1 regularization on the projection coefficients when using iterative operators to obtain NMF basis vectors for the effective tracking. Secondly, to obtain the sparse error and projection coefficient matrice, we present an iterative algorithm to solve the optimal problem, which ensures the representation is more robust. Finally, we take partial occlusion into construct likelihood function, and combined with INMF learning to update appearance model for alleviating tracking drift. Experimental results compared with the state-of-the-art tracking methods demonstrate the proposed algorithm achieves favorable performance when the object undergoes large occlusion, motion blur and illumination changes.

Adaptive multifeature visual tracking in a probability-hypothesis-density filtering framework

Probability hypothesis density (PHD) based trackers have enjoyed growing popularity in recent years, particularly in the field of nonlinear non-Gaussian visual tracking scenarios. These visual trackers can be degraded by a variety of factors, including changes of illumination, occlusion, poor image contrast, shape and appearance variation, clutter and other unmodeled changes of tracked objects. In this paper, for enhancing the performance of PHD based trackers, both scale invariant feature and color distribution feature are used as descriptors of targets of interest. To adaptively adjust the weights of each feature in the tracking process, a confidence measure, i.e., a quantitative measure for the spatial uncertainty of each feature is incorporated into the multifeature tracking algorithm. Experimental results show that the proposed multifeature tracker can improve the reliability and robustness of state estimation and the number estimation in tracking a variable number of objects of varying scales even when background region was similar to the objects appearance.

Probability-hypothesis-density filter for multitarget visual tracking with trajectory recognition

The probability-hypothesis-density (PHD) filter as a multitarget recursive Bayes filter has generated substantial interest in the visual tracking field due to its ability to handle a time-varying number of targets. But the targets trajectory cannot be identified within its own framework. To complement the ability of PHD, the auction algorithm is combined to calculate the object trajectories automatically. We present a motion detection, dynamic, and measurement equation, as well as visual multitarget tracking algorithm based on Gaussian mixture probability hypothesis density with trajectory computation in detail. Experimental results on a large video surveillance dataset show that the proposed multitarget tracking framework improves the tracker and recognizes tracks when a variable number of targets appear, merge, split, and disappear, even in cluttered scenes.

Anomaly Detection Based on Local Nearest Neighbor Distance Descriptor in Crowded Scenes

We propose a novel local nearest neighbor distance (LNND) descriptor for anomaly detection in crowded scenes. Comparing with the commonly used low-level feature descriptors in previous works, LNND descriptor has two major advantages. First, LNND descriptor efficiently incorporates spatial and temporal contextual information around the video event that is important for detecting anomalous interaction among multiple events, while most existing feature descriptors only contain the information of single event. Second, LNND descriptor is a compact representation and its dimensionality is typically much lower than the low-level feature descriptor. Therefore, not only the computation time and storage requirement can be accordingly saved by using LNND descriptor for the anomaly detection method with offline training fashion, but also the negative aspects caused by using high-dimensional feature descriptor can be avoided. We validate the effectiveness of LNND descriptor by conducting extensive experiments on different benchmark datasets. Experimental results show the promising performance of LNND-based method against the state-of-the-art methods. It is worthwhile to notice that the LNND-based approach requires less intermediate processing steps without any subsequent processing such as smoothing but achieves comparable event better performance.

New method for dynamic bias estimation: Gaussian mean shift registration

A novel algorithm, Gaussian mean shift registration (GMSR), is proposed for multisensor dynamic bias estimation. The sufficient condition for convergence of a Gaussian mean shift procedure is given, which extends the current theorem from a strictly convex kernel to a piece-wise convex and concave kernel. The Gaussian mean shift algorithm combined with the extended Kalman filter (EKF) is implemented to estimate the dynamic bias based on the measurements from a single target, which is an iterative optimization procedure. Monte Carlo simulations show that the new algorithm has significant improvement in performance with reducing root mean square (RMS) errors compared with the minimum mean square error (MMSE) estimator, based on multiple targets and multiple frames. The proposed estimator is close to the theoretical lower bound, i.e., it is more efficient in estimating the dynamic bias than other methods.

Graphic-processing-unit-accelerated real-time exposure fusion method using pixel-level optimal exposure criterion

High dynamic range (HDR) imaging is an important and challenging research topic in computational photography. A simple but effective image fusion method is proposed to accomplish the multi-exposure image composition in both static and dynamic scenes. The foundation of the proposed method is an experiential criterion that optimizes the exposure that occurs at a dramatic alteration point in the low dynamic range image sequence (LDRI). To extract these well-exposed pixel vectors, each pixel curve formed by the pixel vectors at same position along all frames in the LDRIs is first preprocessed by the chord length parameterization. Then a single high-quality pseudo-HDR image can be extracted directly and efficiently from the LDRIs using a pixel-level fusion index matrix derived from the first- and second-order difference quotients of the preprocessed pixel curves. The main advantage of the proposed method is its use of a single independent pixel in computing. It is highly parallel, allowing a graphic processing unit-based, real-time implementation. The experiments on various scenes discussed here indicate that the proposed exposure fusion method can combine a large image sequence with 10 megapixels into a visually compelling pseudo-HDR image at a rate of 30  frames/s on a consumer hardware.

Face Classification Based on Natural Features and Decision Tree

Existing face recognition methods suffer from efficiency problems and heavily rely on proper features extraction. In this paper, we propose an efficient face classification method which aims to reduce sensitivity to facial variations and occlusions, meanwhile complete tasks efficiently. In contrast with most energy minimizing based recognition methods, proposed algorithm is cast as a simple classification in our method. First, preprocess images for enhancing data images prior to computational processing and label parts of images as training data. Then we use Active Shape Model (ASM) to extract robust natural features. After that we categorize features and mark them with different labels. Finally, we learn a discriminative C4.5 decision tree for classification. Our method can efficiently classify face images and robust handle facial variations and occlusions. Extensive experiments are conducted on AR database in order to demonstrate the robustness of proposed method. Quantitative and qualitative results compared with several popular algorithms suggest effectiveness and efficiency of proposed method.

On the sensor order in sequential integrated probability data association filter

The processing order of sensors with different detection probabilities and in different clutter densities in a multi-sensor system is investigated in this paper. A sequential implementation of the integrated probability data association (IPDA) algorithm under random set framework is derived. Under the assumptions of different detection probabilities and different clutter densities of individual sensor in a multi-sensor system, we reach the conclusion that the sequential IPDA filter depends on the order analyzing the target existence probability of varying sensor orders. Moreover, we obtain the optimal order of sensors for the sequential IPDA filter in terms of maximizing the target existence probability. The conclusions are demonstrated by simulation results.

PHD filter for multi-target visual tracking with trajectory recognition

Probability hypothesis density (PHD) filter, as a multi-target recursive Bayes filter, has generated substantial interest in the visual tracking field due to its ability to handle a time-varying number of nonlinear targets. But the targets trajectory cannot be identified within its own framework. To complement the ability of PHD, the auction algorithm is combined to calculate the object trajectories automatically. We present the motion detection, dynamic and measurement equation, as well as visual multi-target tracking algorithm based on Gaussian mixture probability hypothesis density (GM-PHD) in details. Experimental results on a large video surveillance dataset show the proposed multi-target tracking framework improves the tracker and recognizes the tracks when a variable number of targets appear, merge, split and disappear even in cluttered scenes.