Yanyun Cheng

Temporal-spatial coherence based abnormal behavior detection

To improve the accuracy and speed of the local abnormal detection, a novel method based on Temporal-Spatial Coherence model is proposed. Specifically, the video block is firstly extracted using the gradient histogram and optimized based on the temporal-spatial coherence. Then the normal behavior model and abnormal behavior model is learned via the tensor voting algorithm and the temporal-spatial coherence respectively. Finally, abnormal behavior is detected and labeled. The experiments conducted on the public UCSD and Subway datasets demonstrate the efficiency of the proposed method for local abnormal behavior detection.

Object detection via superpixel and 3D self-organizing background subtraction

For moving targets with slow speed and temporary stationary, the detection performance of traditional methods via Gauss model and three-frame model is not so satisfactory. Therefore, a novel scheme is proposed to improve the detection performance. Specifically, the simple linear iterative clustering algorithm method is first utilized to complete the superpixel segmentation; then, the 3D self-organizing background subtraction algorithm is utilized to achieve the background model; finally, the optimal weight decision strategy is utilized to detect moving targets. Experimental results conducted on MSA and PETS2009 datasets demonstrate that the proposed scheme can effectively improve the object detection performance.

Image search via semantic hashing learning

With the rapid proliferation of large-scale web images, recent years have witnessed more and more images labeled with user-provided tags, which leads to considerable effort made on hashing based image retrieval in huge databases. Current research efforts focus mostly on learning semantic hashing functions which designs compact binary codes to map semantically similar images to similar codes, and the visual similarity is not well explored for constructing semantic hashing functions. Here a novel approach is proposed to learn hashing functions that preserve semantic and visual similarity between images. Specifically, semantic hashing codes are first learned by leveraging the similarity between textual structure and visual structure; then, maximum entropy principle is exploited to achieve compact binary codes; finally, function decay principle is introduced to remove noisy visual attributes. Experimental results conducted on a widely-used image dataset demonstrate the proposed approach can effectively improve the performance in image retrieval.

Motion coherence based abnormal behavior detection

To improve the accuracy and speed of the global abnormal detection, a novel method based on motion coherence model is here proposed. Specifically, motion features of each tracking objects are firstly extracted; then, global abnormal behavior detection models are learned based on the energy model, dispersion model and Lagrange particle dynamics model respectively; finally, global abnormal behavior is detected and labeled based on the learned three models. The proposed method is conducted on public UMN dataset which demonstrates that the proposed method can improve the accuracy and efficiency of abnormal behavior detection.

Object detection via optical information and background information

This paper presents a novel method to detect salient objects by exploiting the optical information and background information. Specifically, each video image is first segmented into superpixels; then, the optical information and color information is utilized to obtain the initial result of each salient object; finally, the accurate result of each salient object is obtained by taking into consideration the background information. Experimental results conducted on the SegTrackv2 database demonstrate the the effectiveness of the proposed method.

Pedestrians tracking based on least squares algorithm and intelligent collision avoidance model

Target tracking is a hot topic in the field of computer vision and pattern recognition. The aim of target tracking is to achieve the location of moving targets and tracking trajectories of moving targets. A pedestrian tracking method based on the Least Squares algorithm and intelligent collision avoidance model is proposed in this paper. Specifically, the traditional Kalman algorithm is first utilized to realize the initial target tracking; then, to deal with the issue of target tracking caused by the traditional Kalman algorithm, the least square method is here utilized to fit the pedestrian moving curve and predict the location of the pedestrians in the next frame, which can be utilized as the initial moving object for the later search; finally, the intelligent collision avoidance algorithm is here proposed to improve the tracking accuracy in case of obstacles.

Salient object detection via harris corner

In this paper, an fast and effective image saliency detection based on Harris Corner method is proposed. Different from most previous methods that mainly concentrate on boundary prior, we take both background and foreground information into consideration. First, a novel method is proposed to approximately locate the foreground object by using the convex hull from Harris corner. Then, the original image is segmented into super-pixels regions and the saliency values of different regions are divided into two parts to generate the corresponding background and foreground cue maps which are combined into a unified map. Finally, the unified map and the convex hull center-biased algorithm are combined to be the saliency map, which is then optimized by Bayesian perspective and saliency diffusion to get the final result. Experiments on publicly available data sets demonstrate that the proposed algorithm performs favorably against the state-of-the-art methods.

Trajectory data association based multiobject tracking

A multitarget tracking method based on data association and trajectory evaluation is proposed in this paper. Firstly, the data association and trajectory evaluation are integrated into the same conditional random field model, which is transformed to obtain the minimum energy problem. Secondly, from the time, if two of the same track label appear with a physical space, the symbiotic label cost is used to constrain the correlation; from the space, the pairwise energy term between the different observation targets is introduced to prevent the occurrence of false data label. Finally, in the discrete continuous optimization process of the energy, this paper uses the improved a-expansion algorithm and the gradient descent method to solve the minimum energy in non convex and non sub module functions. The experimental results of PETS2009/2010 benchmark and TUD-Stadtmitte video sequence database show that the proposed algorithm in this paper is superior to the current advanced level of multitarget tracking technology.

Image classification via multi-view model

With the massive growth of digital image data uploaded to the Internet, classifying each image into appropriate semantic category based on the image content for image index and image retrieval has become an increasingly difficult and laborious task. To deal with this issue, we propose a novel multi-view semi-supervised learning framework which leverages the information contained in pseudo-labeled images to improve the prediction performance of image classification using multiple views of an image. In the training process, labeled images are first adopted to train view-specific classifiers independently using uncorrelated and sufficient views, and each view-specific classifier is then interactively re-trained using initial labeled samples and additional pseudo-labeled samples based on a measure of confidence. In the classification process, the maximum entropy principle is utilized to assign appropriate category label to each unlabeled image using optimally trained view-specific classifiers. Experimental results on a general-purpose image database demonstrate the effectiveness and efficiency of the proposed multi-view semi-supervised scheme.

Improved balanced binary tree based action recognition

Action recognition is one of the core content of intelligent monitoring, and also the basis of video content analysis and understanding. A novel method is here proposed to enhance the accuracy of human behavior recognition. First, each video image is divided into five sub-regions based on the motion mechanism; then, the frequency information of optical flow within each sub-region is extracted to describe the motion characteristics of each sub-region; finally, an improved balanced binary decision tree-support vector machine is utilized to complete the task of behavior recognition. Experimental results conducted on KTH database demonstrate the proposed algorithm can improve the accuracy of behavior recognition.