Ching-Tang Fan

Single image defogging by multiscale depth fusion.

Restoration of fog images is important for the deweathering issue in computer vision. The problem is ill-posed and can be regularized within a Bayesian context using a probabilistic fusion model. This paper presents a multiscale depth fusion (MDF) method for defog from a single image. A linear model representing the stochastic residual of nonlinear filtering is first proposed. Multiscale filtering results are probabilistically blended into a fused depth map based on the model. The fusion is formulated as an energy minimization problem that incorporates spatial Markov dependence. An inhomogeneous Laplacian–Markov random field for the multiscale fusion regularized with smoothing and edge-preserving constraints is developed. A nonconvex potential, adaptive truncated Laplacian, is devised to account for spatially variant characteristics such as edge and depth discontinuity. Defog is solved by an alternate optimization algorithm searching for solutions of depth map by minimizing the nonconvex potential in the random field. The MDF method is experimentally verified by real-world fog images including cluttered-depth scene that is challenging for defogging at finer details. The fog-free images are restored with improving contrast and vivid colors but without over-saturation. Quantitative assessment of image quality is applied to compare various defog methods. Experimental results demonstrate that the accurate estimation of depth map by the proposed edge-preserved multiscale fusion should recover high-quality images with sharp details.

Real-time camera anomaly detection for real-world video surveillance

This paper proposes an automatic event detection technique for camera anomaly by image analysis, in order to confirm good image quality and correct field of view of surveillance videos. The technique first extracts reduced-reference features from multiple regions in the surveillance image, and then detects anomaly events by analyzing variation of features when image quality decreases and field of view changes. Event detection is achieved by statistically calculating accumulated variations along temporal domain. False alarms occurred due to noise are further reduced by an online Kalman filter that can recursively smooth the features. Experiments are conducted on a set of recorded videos simulating various challenging situations. Compared with an existing method, experimental results demonstrate that our method has high precision and low false alarm rate with low time complexity.

Heterogeneous Information Fusion and Visualization for a Large-Scale Intelligent Video Surveillance System

Wide-area monitoring for a smart community can be challenging in systems engineering because of its large scale and heterogeneity at the sensor, algorithm, and visualization levels. A smart interface to visualize high-level information fused from a diversity of low-level surveillance data, and to facilitate rapid response of events, is critical for the design of the system. This paper presents an event-driven visualization mechanism fusing multimodal information for a large-scale intelligent video surveillance system. The mechanism proactively helps security personnel intuitively be aware of events through close cooperation among visualization, data fusion, and sensor tasking. The visualization not only displays 2-D, 3-D, and geographical information within a condensed form of interface but also automatically shows the only important video streams corresponding to spontaneous alerts and events by a decision process called display switching arbitration. The display switching arbitration decides the importance of cameras by score ranking that considers event urgency and semantic object features. This system has been successfully deployed in a campus to demonstrate its usability and efficiency for an installation with two camera clusters that include dozens of cameras, and with a lot of video analytics to detect alerts and events. A further simulation comparing the display switching arbitration with similar camera selection methods shows that our method improves the visualization by selecting better representative camera views and reducing redundant switchover among multiview videos.

Accurate depth estimation for image defogging using Markov random field

This paper presents an automatic method for the defogging process from a single haze image. To recover a foggy image, an accurate depth map is estimated from a multi-level estimation method, which fuses depth maps with different sizes of patches by dark channel prior. Markov random field (MRF) is applied to label the depth level in adjacent region for the compensation of wrong estimated regions. The accurate estimation of scene depth provides good restoration with respect to visibility and contrast but without oversaturating. The algorithm is verified by a handful of foggy and hazy images. Experimental results demonstrate that the defogging method can recover high-quality images through accurate estimation of depth map.

Traffic Camera Anomaly Detection

Detection of camera anomaly and tampering have attracted increasing interest in video surveillance for real-time alert of camera malfunction. However, the anomaly detection for traffic cameras monitoring vehicles and recognizing license plates has not been formally studied and it cannot be solved by existing methods. In this paper, we propose a camera anomaly detection method for traffic scene that has distinct characteristics of dynamics due to traffic flow and traffic crowd, compared with normal surveillance scene. Image quality used as low-level features are measured by no-referenced metrics. Image dynamics used as mid-level features are computed by histogram distribution of optical flow. A two-stage classifier for the detection of anomaly is devised by the modeling of image quality and video dynamics with probabilistic state transition. The proposed approach is robust to many challenging issues in urban surveillance scenarios and has very low false alarm rate. Experiments are conducted on real-world videos recorded in traffic scene including the situations of high traffic flow and severe crowding. Our test results demonstrate that the proposed method is superior to previous methods on both precision rate and false alarm rate for the anomaly detection of traffic cameras.

Home sleep care with video analysis and its application in smart TV

In this paper, we present a home sleep care system to monitor function of activity measurement and sleeping position estimation via vision-based algorithms. The outcomes are integrated in a telecare middleware with a hierarchical software stack and visualized by a structural subsystem. Since smart TVs play an important role in home tele-healthcare system, we centralize and display additional sleep care information at a client/server system. The proposed method has been verified to be highly correlated to the well-established sleep monitoring device, actigraphy. In addition, an innovative prototype demonstration with sleep care has been revealed to apply in smart TV environments.

Moving Object Detection for Night Surveillance

Traditional background subtraction methods perform poorly at night. In this paper, a robust method is proposed for automatic visual surveillance in low-light level environment which has quality problems of low brightness, low contrast and high-level noise. The novel method includes techniques of illumination compensation and illumination-invariant background subtraction to solve the low-quality problem in night surveillance. Experiments are conducted on several challenging videos captured with drastic illumination change at night. Experimental results demonstrate that the proposed approach significantly outperforms existing techniques for the extraction of moving objects at night.

A Large Scale Video Surveillance System with Heterogeneous Information Fusion and Visualization for Wide Area Monitoring

Wide area monitoring for community and city can be a very challenging engineering task due to its scale and heterogeneity in sensor, algorithm, and visualization levels. Multi-modal cameras and algorithms have to be fused into compact presentation for a single operator to actively and effectively respond to anomaly events and jeopardy. This paper presents a distributed and scalable video surveillance system, subcontracted by intelligent surveillance components (ISCs) and visualization surveillance components (VSCs) in compliance with functional labors. The ISCs are high-level algorithms applying computer vision for behavioral analysis of human and vehicles. The VSCs constitute a multi-tier subsystem to visualize fused results of messages, key frames, streaming videos and geographic context information. The system helps the operator to focus attention on interested events gathered from distrusted ISCs and presented by VSCs on map and three-dimensional homographic views. Robustness and effectiveness of the system has been demonstrated by a test run of real scenarios deployed in a campus.

Accelerating multi-scale retinex using ARM NEON

High dynamic range image processing have recently become an important topic in consumer electronics market. While multi-scale retinex with color restoration (MSRCR) have been well developed, disadvantages of low performance is not favorable to a mobile computer-vision system. To remedy the above problem, this paper proposes an accelerated MSRCR with effective use of ARM Cortex-A9 architecture and NEON SIMD technology. A linear sampling method with binomial normal approximation is developed for improving performance of Gaussian smoothing. Overall performance improvement of MSRCR algorithm on Zedboard platform is 74% compared to original ARM optimized C code compiled to Cortex-A9 processor architecture.

Multiscale fusion of depth estimations for haze removal

Restoration of haze images is important for the de-weathering issue in computer vision. The problem is ill-posed and can be regularized within a Bayesian context by using a probabilistic fusion model. This paper presents a multiscale depth fusion (MDF) method for dehazing from a single image. A linear model representing the stochastic residual of nonlinear filtering is first proposed. Multiscale filtering results are probabilistically blended into a fused depth map based on the model. The fusion is formulated as an energy minimization problem that incorporates spatial Markov dependency. An inhomogeneous Laplacian-Markov random field for the multiscale fusion regularized with smoothing and edge-preserving constraints is developed. The MDF method is experimentally verified by cluttered-depth image that is challenging for dehaze at finer details. Experimental results demonstrate that the accurate estimation of depth map by the proposed edge-preserved multiscale fusion should recover high-quality images with sharp details.