Yiping Yang

Ship detection by salient convex boundaries

Automatic ship detection from remote sensing imagery has many applications, such as maritime security, traffic surveillance, fisheries management. However, it is still a difficult task for noise and distractors. This paper is concerned with perceptual organization, which detect salient convex structures of ships from noisy images. Because the line segments of contour of ships compose a convex set, a local gradient analysis is adopted to filter out the edges which are not on the contour as preprocess. For convexity is the significant feature, we apply the salience as the prior probability to detect. Feature angle constraint helps us compute probability estimate and choose correct contour in many candidate closed line groups. Finally, the experimental results are demonstrated on the satellite imagery from Google earth.

Visual Tracking via Sparse Representation Based Linear Subspace Model

The modeling of the object appearance is one of the key issues in the development and application of effective object tracking. This paper presents a tracking algorithm based on representing the appearance of the object using a sparse representation based subspace model. the sparse representation theory offers us a powerful tool to model the object by only a small fraction of the training set. The multi-part subspace appearance model(MSAM) is learned via L1-minimization and the Gramm-Schmidt process given enough training samples (overcomplete dictionary). Furthermore, a novel model updating strategy is designed to incrementally update the proposed subspace model and the dictionary. Finally, an observation model integrating both sparsity and the likelihood information is designed to embed the proposed modeling approach into the particle filter framework for efficient object tracking. Experimental results demonstrate the robustness and effectiveness of the algorithm, especially when the images are noisy or the objects exhibit large appearance changes.

Anomaly detection in hyperspectral imagery based on low-rank and sparse decomposition

This paper presents a novel low-rank and sparse decomposition (LSD) based model for anomaly detection in hyperspectral images. In our model, a local image region is represented as a low-rank matrix plus spares noises in the spectral space, where the background can be explained by the low-rank matrix, and the anomalies are indicated by the sparse noises. The detection of anomalies in local image regions is formulated as a constrained LSD problem, which can be solved efficiently and robustly with a modified “Go Decomposition” (GoDec) method. To enhance the validity of this model, we adapts a “simple linear iterative clustering” (SLIC) superpixel algorithm to efficiently generate homogeneous local image regions i.e. superpixels in hyperspectral imagery, thus ensures that the background in local image regions satisfies the condition of low-rank. Experimental results on real hyperspectral data demonstrate that, compared with several known local detectors including RX detector, kernel RX detector, and SVDD detector, the proposed model can comfortably achieves better performance in satisfactory computation time.

Parallel of low-level computer vision algorithms on a multi-DSP system

Parallel hardware becomes a commonly used approach to satisfy the intensive computation demands of computer vision systems. A multiprocessor architecture based on hypercube interconnecting digital signal processors (DSPs) is described to exploit the temporal and spatial parallelism. This paper presents a parallel implementation of low level vision algorithms designed on multi-DSP system. The convolution operation has been parallelized by using redundant boundary partitioning. Performance of the parallel convolution operation is investigated by varying the image size, mask size and the number of processors. Experimental results show that the speedup is close to the ideal value. However, it can be found that the loading imbalance of processor can significantly affect the computation time and speedup of the multi- DSP system.

Adaptive dynamic model particle filter for visual object tracking

One of the key issues related to object tracking is the representation of the object motion. It is a challenging problem because the object usually exhibits complex and rich dynamic behavior. In this paper, we propose an adaptive dynamic model to describe the dynamics/motion of the object and embed it into the particle filter framework for visual object tracking. The model characterize the object motion preciously by a switch-and-fusion strategy, which integrates both long period and short period motion information by the combination of multiple simple motion models. Experimental results demonstrate that, the proposed method achieves better results than the conventional particle filter, especially when the object moves quickly and changes the motion pattern drastically.

A Novel Method for Exactly Locating Road Intersections in Aerial Images

Road intersections in aerial images could provide lots of geography information, and thus are very valuable fornavigation systems, e.g. cruise missile or UAV(unmanned Aerial Vehicle) navigation system. However, there are few algorithms presented to exactly locate road intersections, and mostly the computational costs of them are very expensive, which can not meet the on-line processing requirement. In this paper, we present a method for exactly locating road intersections. Firstly, we limit the candidate area around the road region, and employ the FAST(Features from Accelerated Segment Test) corner detector to obtain some candidate road intersections. Then, a novel model is proposed to confirm and classify the real road intersections. Finally, the MS(Mean-Shift) algorithm isused here to move them to their most proper positions. Several aerial images are taken to test the effectiveness of our method.Experimental results demonstrated that the correctness andveracity are both acceptable, and the cost of our method isvery low.

Predicting Chinese Stock Market Price Trend Using Machine Learning Approach

The stock1 market is dynamic, noisy and hard to predict. In this paper, we explored four machine learning models using technical indicators as input features to predict the price trend 30 days later. The experimental dataset is Shanghai Stock Exchange(SSE) 50 index stocks. The result demonstrates that ANN performs better than the other three models and is promising to find some profitable patterns.

Communication in a hybrid multi-layer MIMD system for computer vision

Multi-layer MIMD computers are variable solutions for computer vision due to their flexibility and computational power. However, performance of parallel programs executed on these systems is dependent on the efficiency of inter-processor communication. Based on a hybrid multi-layer MIMD computer vision computer and message-passing mechanism, we present a software system developed to handle complex communication operations. Communication model, routing algorithm and switching technique are established and evaluated. Extensive experiments have been conducted to determine the hardware parameters of the communication links of the computer. Furthermore, communication algorithms that are frequently required in parallel vision algorithms are developed and their performance is presented.

Self-Recalibration of PTZ Cameras

Pan-tilt-zoom (PTZ) cameras have been widely used for providing flexible view selection and a wider observation range. In order to utilize these cameras for a meaning application, it is necessary to calibrate cameras. However, with camera pan, tilt and zoom, the previous calibrated camera parameters often vary, which are required to be updated automatically. In this paper, we present a novel self-recalibtion method of PTZ cameras under a small pitch angle. The proposed approach is based on the calibrated camera. First, the keypoints from two frames, corresponding to ones before and after the parameters are changed, are matched roughly, then the false matching pairs are removed by the displacement constraint, which is statistical displacement constraint in pan and tilt, while in zoom is principal point constraint. Second, the homography between the two frames can be estimated from the matching ones, then the camera parameters can be updated by combining previous known parameters. Furthermore, we present a evaluation method of recalibration precision. Experimental results demonstrate the effectiveness of the proposed algorithm.

On-line feature enhancement for adaptive object tracking

This paper presents an adaptive tracking algorithm by online features enhancement. To avoid the distraction of the similar background on tracker, Bayes decision rule is applied to calculate the posterior probability of every pixel belonging to the object and generate a set of candidate confidence maps according to the conditional sample densities from object and background on different features. We evaluate the performance of every candidate confidence map using moment of inertia. Then, an optimal confidence map is selected to be fed to Meanshift which is employed to find the location of the object. At last, we update the target model by the confidence map. Experimental validation of the proposed method is performed and presented on challenging image sequences.