A novel approach for detecting anomaly in visual surveillance system is proposed in this paper. It is composed of three parts: (a) a dense motion field and motion statistics method, (b) motion directional PCA for feature dimensionality reduction, (c) an improved one-class SVM for one-class classification. Experiments demonstrate the effectiveness of the proposed algorithm in detecting abnormal events in surveillance video, while keeping a low false alarm rate. Our scheme works well in complicated situations that common tracking or detection modules cannot handle.

Boosting over weak classifiers is widely used in pedestrian detection. As the number of weak classifiers is large, researchers always use a sampling method over weak classifiers before training. The sampling makes the boosting process harder to reach the fixed target. In this paper, we propose a partial derivative guidance for weak classifier mining method which can be used in conjunction with a boosting algorithm. Using weak classifier mining method makes the sampling less degraded in the performance. It has the same effect as testing more weak classifiers while using acceptable time. Experiments demonstrate that our algorithm can process quicker than [1] algorithm in both training and testing, without any performance decrease. The proposed algorithms is easily extending to any other boosting algorithms using a window-scanning style and HOG-like features.

In this letter, we present a flow-aware opportunistic routing protocol over wireless mesh networks. Firstly, a forwarder set selection mechanism is proposed to avoid potential flow contention, thus alleviating possible congestion from the increased number of flows. Secondly, a Round-Robin packet sending fashion combined with batch-by-batch acknowledgement is introduced to provide reliability and improve throughput. Evaluations show that our protocol significantly outperforms a seminal opportunistic routing protocol, MORE, under both single and multiple flow scenarios.

Object tracking is a major technique in image processing and computer vision. Tracking speed will directly determine the quality of applications. This paper presents a parallel implementation for a recently proposed scale- and rotation-invariant on-line object tracking system. The algorithm is based on NVIDIA's Graphics Processing Units (GPU) using Compute Unified Device Architecture (CUDA), following the model of single instruction multiple threads. Specifically, we analyze the original algorithm and propose the GPU-based parallel design. Emphasis is placed on exploiting the data parallelism and memory usage. In addition, we apply optimization technique to maximize the utilization of NVIDIA's GPU and reduce the data transfer time. Experimental results show that our GPGPU-based method running on a GTX480 graphics card could achieve up to 12X speed-up compared with the efficiency equivalence on an Intel E8400 3.0 GHz CPU, including I/O time.

The time efficiency of aerial video stitching is still an open problem due to the huge amount of input frames, which usually results in prohibitive complexities in both image registration and blending. In this paper, we propose an efficient framework aiming to stitch aerial videos in real time. Reasonable distortions are allowed as a tradeoff for acceleration. Instead of searching for globally optimized solutions, we directly refine frame positions with sensor data to compensate for the accumulative error in alignment. A priority scan method is proposed to select pixels within overlapping area into the final panorama for blending, which avoids complicated operations like weighting or averaging on pixels. Experiments show that our method can generate satisfying results at very competitive speed.

Image sequence registration has attracted increasing attention due to its significance in image processing and computer vision. In this paper, we put forward a new kernel based image registration approach, combining both feature-based and intensity-based methods. The proposed algorithm consists of two steps. The first step utilizes feature points to roughly estimate a motion parameter between successive frames; the second step applies our kernel based idea to align all the frames to the reference frame (typically the first frame). Experimental results using both synthetic and real image sequences demonstrate that our approach can automatically register all the image frames and be robust against illumination change, occlusion and image noise.

Based on sample-pair refinement and local optimization, this paper proposes a high-accuracy and quick matting algorithm. First, in order to gather foreground/background samples effectively, we shoot rays in hybrid (gradient and uniform) directions. This strategy utilizes the prior knowledge to adjust the directions for effective searching. Second, we refine sample-pairs of pixels by taking into account neighbors'. Both high confidence sample-pairs and usable foreground/background components are utilized and thus more accurate and smoother matting results are achieved. Third, to reduce the computational cost of sample-pair selection in coarse matting, this paper proposes an adaptive sample clustering approach. Most redundant samples are eliminated adaptively, where the computational cost decreases significantly. Finally, we convert fine matting into a de-noising problem, which is optimized by minimizing the observation and state errors iteratively and locally. This leads to less space and time complexity compared with global optimization. Experiments demonstrate that we outperform other state-of-the-art methods in local matting both on accuracy and efficiency.

Person re-identification is challenging due to illumination changes and viewpoint variations in the multi-camera environment. In this paper, we propose a novel spatial pyramid color representation (SPCR) and a local region matching scheme, to explore person appearance for re-identification. SPCR effectively integrates color layout into histogram, forming an informative global feature. Local region matching utilizes region statistics, which is described by covariance feature, to find appearance correspondence locally. Our approach shows robustness to illumination changes and slight viewpoint variations. Experiments on a public dataset demonstrate the performance superiority of our proposal over state-of-the-art methods.

This paper presents a self-clustering algorithm to detect symmetry in images. We combine correlations of orientations, scales and descriptors as a triple feature vector to evaluate each feature pair while low confidence pairs are regarded as outliers and removed. Additionally, all confident pairs are preserved to extract potential symmetries since one feature point may be shared by different pairs. Further, each feature pair forms one cluster and is merged and split iteratively based on the continuity in the Cartesian and concentration in the polar coordinates. Pseudo symmetric axes and outlier midpoints are eliminated during the process. Experiments demonstrate the robustness and accuracy of our algorithm visually and quantitatively.

Robust local image features have become crucial components of many state-of-the-art computer vision algorithms. Due to limited hardware resources, computing local features on embedded system is not an easy task. In this paper, we propose an efficient parallel computing framework for speeded-up robust features with an orientation towards multi-DSP based embedded system. We optimize modules in SURF to better utilize the capability of DSP chips. We also design a compact data layout to adapt to the limited memory resource and to increase data access bandwidth. A data-driven barrier and workload balance schemes are presented to synchronize parallel working chips and reduce overall cost. The experiment shows our implementation achieves competitive time efficiency compared with related works.

Particle filter has attracted increasing attention from researchers of object tracking due to its promising property of handling nonlinear and non-Gaussian systems. In this paper, we mainly explore the problem of precisely estimating observation likelihoods of particles in the joint feature-spatial space. For this purpose, a mixture Gaussian kernel function based similarity is presented to evaluate the discrepancy between the target region and the particle region. Such a similarity can be interpreted as the expectation of the spatial weighted feature distribution over the target region. To adapt outburst of object motion, we also present a method to appropriately adjust state transition model by utilizing the priors of motion speed and object size. In comparison with the standard particle filter tracker, our tracking algorithm shows the better performance on challenging video sequences.

This paper addresses stereo matching under scenarios of smooth region and obviously slant plane. We explore the flexible handling of color disparity, spatial relation and the reliability of matching pixels in support windows. Building upon these key ingredients, a robust stereo matching algorithm using local plane fitting by Confidence-based Support Window (CSW) is presented. For each CSW, only these pixels with high confidence are employed to estimate optimal disparity plane. Considering that RANSAC has shown to be robust in suppressing the disturbance resulting from outliers, we employ it to solve local plane fitting problem. Compared with the state of the art local methods in the computer vision community, our approach achieves the better performance and time efficiency on the Middlebury benchmark.

Learning an appearance model for person re-identification from multiple images is challenging due to the corrupted images caused by occlusion or false detection. Furthermore, different persons may wear similar clothes, making appearance feature less discriminative. In this paper, we first introduce the concept of multiple instance to handle corrupted images. Then a novel pairwise comparison based multiple instance learning framework is proposed to deal with visual ambiguity, by selecting robust features through pairwise comparison. We demonstrate the effectiveness of our method on two public datasets.

The transfer of prior knowledge from source domains can improve the performance of learning when the training data in a target domain are insufficient. In this paper we propose a new strategy to transfer deformable part models (DPMs) for object detection, using offline-trained auxiliary DPMs of similar categories as source models to improve the performance of the target object detector. A DPM presents an object by using a root filter and several part filters. We use these filters of the auxiliary DPMs as prior knowledge and adapt the filters to the target object. With a latent transfer learning method, appropriate local features are extracted for the transfer of part filters. Our experiments demonstrate that this strategy can lead to a detector superior to some state-of-the-art methods.

This work presents a real-time human detection system for VGA (Video Graphics Array, 640480) video, which well suits visual surveillance applications. To achieve high running speed and accuracy, firstly we design multiple fast scalar feature types on the gradient channels, and experimentally identify that NOGCF (Normalized Oriented Gradient Channel Feature) has better performance with Gentle AdaBoost in cascaded classifiers. A confidence measure for cascaded classifiers is developed and utilized in the subsequent tracking stage. Secondly, we propose to use speedup techniques including a detector pyramid for multi-scale detection and channel compression for integral channel calculation respectively. Thirdly, by integrating the detector's discrete detected humans and continuous detection confidence map, we employ a two-layer tracking by detection algorithm for further speedup and accuracy improvement. Compared with other methods, experiments show the system is significantly faster with 20 fps running speed in VGA video and has better accuracy as well.

Human detection has witnessed significant development in recent years. The introduction of cascade structure and integral histogram has greatly improved detection speed. But real-time detection is still only possible for sparse scan of 320 240 sized images. In this work, we propose a matrix-based structure to reorganize the computation structure of window-scanning detection algorithms, as well as a new pre-processing method called Hierarchical HOG Matrices (HHM) in place of integral histogram. Our speed-up scheme can process 320 240 sized images by dense scan (≈ 12000 windows per image) at the speed of about 30 fps, while maintaining accuracy comparable to the original HOG + cascade method.

This paper presents a robust framework of human-computer interaction from the hand gesture vision in the presence of realistic and challenging scenarios. To this end, several novel components are proposed. A hybrid approach is first proposed to automatically infer the beginning position of hand gestures of interest via jointly optimizing the regions given by an offline skin model trained from Gaussian mixture models and a specific hand gesture classifier trained from the Adaboost technique. To consistently track the hand in the context of using kernel based tracking, a semi-supervised feature selection strategy is further presented to choose the feature subspaces which appropriately represent the properties of offline hand skin cues and online foreground-background-classification cues. Taking the histogram of oriented gradients as the descriptor to represent hand gestures, a soft-decision approach is finally proposed for recognizing static hand gestures at the locations where severe ambiguity occurs and hidden Markov model based dynamic gestures are employed for interaction. Experiments on various real video sequences show the superior performance of the proposed components. In addition, the whole framework is applicable to real-time applications on general computing platforms.

In this paper, we propose an interleaving updating framework of disparity and confidence map (IUFDCM) for stereo matching to eliminate the redundant and interfere information from unreliable pixels. Compared with other propagation algorithms using matching cost as messages, IUFDCM updates the disparity map and the confidence map in an interleaving manner instead. Based on the Confidence-based Support Window (CSW), disparity map is updated adaptively to alleviate the effect of input parameters. The reassignment for unreliable pixels with larger probability keeps ground truth depending on reliable messages. Consequently, the confidence map is updated according to the previous disparity map and the left-right consistency. The top ranks on Middlebury benchmark corresponding to different error thresholds demonstrate that our algorithm is competitive with the best stereo matching algorithms at present.

Various observation models have been introduced into the object tracking community, and combining them has become a promising direction. This paper proposes a novel approach for estimating the confidences of different observation models, and then effectively combining them in the particle filter framework. In our approach, spatial Likelihood distribution is represented by three simple but efficient parameters, reflecting the overall similarity, distribution sharpness and degree of multi peak. The balance of these three aspects leads to good estimation of confidences, which helps maintain the advantages of each observation model and further increases robustness to partial occlusion. Experiments on challenging video sequences demonstrate the effectiveness of our approach.

This paper proposes a high-accuracy sub-pixel registration framework based on phase correlation for noisy images. First we introduce a denoising module, where the edge-preserving filter is adopted. This strategy not only filters off the noise but also preserves most of the original image signal. A confidence-weighted optimization module is then proposed to fit the linear phase plane discriminately and to achieve sub-pixel shifts. Experiments demonstrate the effectiveness of the combination of our modules and improvements of the accuracy and robustness against noise compared to other sub-pixel phase correlation methods in the Fourier domain.