Demo


In a practical environment, the viewing angle and height of a video surveillance camera are uncontrollable. This may cause severe inter-object occlusion and complicate the detection problem. In this paper, we proposed a novel inference framework with multiple layers forvacantparking space detection. The framework consists of an Image layer, a Patch layer, a Space layer, and a Lot layer. In the Image layer, image patches were selected based on the 3D parking lot structure. We found that the occlusion pattern within each patchrevealscuesof the parking status. Thus, our system extracted lighting-invariant features of patches and trained weak classifiers for the recognition of the occlusion pattern in the Patch layer. The outputs of the classifiers, presenting the types of inter-object occlusion, were treated as the mid-level features and inputted to the Space layer. Next, a boosted space classifier was trained to recognize the mid-level features and output the status of a 3-space unit in a probability fashion. In the Lot layer, we regarded the local status decision of 3-space units as high-level evidences and proposed a Markov Random Field to refine the parking status. In addition, we extended the framework to bridge multiple cameras and integrate the complementary information for vacant space detection.Our results show that the proposed framework can overcome the inter-object occlusion and achieve betterstatus inferencein many environmental variations and different weather conditions.We also presented a real-time system to demonstrate the computing efficiency and the system robustness.


The purpose of this project is to use the image analysis, identification technology, and learning algorithms to analyze abnormal traffic behavior to support the automatic driving system simulation, decision and warning. Real driving behavior is extremely complex, driving behavior often varies from person to person, can not simply classify the behavior model for several simple classification. In order to achieve a systematic analysis, the logical understanding of driving behavior needs to rely on the characteristics of the initial information, the middle of the object information, high-level events to be able to complete the understanding. The key technology includes (1) object detection and image segmentation and (2) according to the object identification detection and image segmentation results, identify the vehicle, pedestrian tracking, and estimate the distance between the scene vehicle and driving.


In this paper, we developed a localization system using smart glasses for museum environments. By taking advantage of the existing QR (Quick Response) codes which are deployed for embedding artwork information in a museum, we proposed a hybrid method using Inertial Measurement Unit (IMU), camera, and Wi-Fi radio signal for visitor localization and tracking. The IMU could provide direct self-motion information for target tracking; however, its measurements might not be robust owing to error propagation from gyro and acceleration biases. Thus, in this paper, a vision system was introduced as a complementary part for localization refinement. Concretely, we utilized smart glasses to detect the existing QR code and based on camera geometry to infer the relative target location by reference to the QR code position. Next, a Kalman Filter (KF) framework was used to derive the coarse location of a visitor by combining the IMU measurements and the vision-based localization result. Finally, the region around the coarse location was treated as a region of interest (ROI) for a Wi-Fi positioning system. Based on the radio signal strength indicator (RSSI) and the trained radio map within the ROI, the visitor location was inferred by applying a K-nearest neighbor (KNN) algorithm. The experiment results show the efficiency of our method compared with an IMU-based method, a Wi-Fi-based method, and an IMU and Wi-Fi fusion method.


Nowadays, there are many service robots in the market; however, only few of them become a popular product. Among them, the vacuum cleaning robot might be the most successful one and treated as the key entry point toward the future market of service robots. In order to enable the intelligent function in a cleaning robot, the ability for a robot to Simultaneous Localization and Mapping (SLAM) is the fundamental and critical step. Hence, in this project, we aim to study and implement the SLAM algorithm in a cleaning robot. Besides, to match the video surveillance purpose of the future product, we plan to adopt an omni-directional camera as an environmental sensor. Our system would use a robot as the research platform with an omni-directional camera embedded on the top. By using the camera to capture omni-directional images, we might have richer information for scene landmark detection and feature extraction. Compared with a standard camera, the function to gather image information from all directions makes an omni-directional camera more suitable for simultaneous localization and mapping. To well use the scene information provided by an omni-directional camera for SLAM, we would spend the major research efforts on the following three technical issues:
(1) Scene landmark detection based on a distorted omni-directional image
(2) Landmark feature description and landmark matching
(3) The integration of SLAM framework and landmark information


In this paper, we discuss a similarity inconsistency phenomenon where the radio signal strength (RSS) signatures of two neighboring positions are dissimilar due to the RSS variation. While matching an observed RSS throughout the radio map, the phenomenon would lead to a jagged similarity distribution. This may break the similarity assumption of the previous works. To address the problem, we proposed a multi-dimensional kernel density estimation (MDKDE) method. By introducing the spatial kernel, the method could adopt neighboring information to enrich the fingerprint. The model can also help to generate a smooth and consistent similarity distribution. Moreover, we formulated the searching of the target location over the continuous domain as an optimization problem. Instead of estimating the optimal location numerically, we also came up with an efficient tracking method, weighted average tracker (WAT). Upon the MDKDE model, WAT can track the target in a simple weighted average method. The experimental results have demonstrated that the proposed system could well model the RSS variation and provide robust positioning performance in an efficient manner.


In this system, we propose a cascaded hierarchical framework for object detection and tracking. We claim that, by integrating both detection and tracking into a unified framework, the detection and tracking of multiple moving objects in a complicated environment become more robust. Under the proposed architecture, detection and tracking cooperate with each other. Based on the result of moving object detection, a dynamic model is adaptively maintained for object tracking. On the other hand, the updated dynamic model is used for both temporal prior propagation of object labels and the update of foreground/background models, which step further to help the detection of moving objects in the next frame. The experiments show that accurate results can be obtained even under situations with camera shaking, foreground/background appearance ambiguity, and object occlusion.


Recently, the computer vision technology for video surveillance applications has made tremendous progress. Those applications may be roughly classified into single-camera systems and multi-camera systems. For a single-camera system, object labeling is an essential step for advanced analysis, like behavior understanding. However, a 2-D image lacks the depth information and thus the detection of moving targets usually suffers from the occlusion problem. The occlusion problem makes it difficult to correctly label or segment connective targets. Moreover, a supervised setting of targets number is usually needed for labeling. Unfortunately, this information is usually not available in practical applications. On the other hand, for a multi-camera system, object correspondence is crucial. The cross reference of multiple camera views may ease the occlusion problem and provide a more reliable way for object labeling. In this system, the major focus is to propose a unified method to label and map targets over multiple cameras. The proposed method can systematically estimate the target number, tackle inter-target occlusions problem, and require neither isolated foreground extraction nor color calibration.


In medical application, X-ray image represents the important matters such as organs, bones, and nodules by the mass over an area unit. The overlap of the matters creates a low contrast and low dynamic range in an X-ray image. Conventional methods relied on ton mapping or filter based algorithms to realize enhancement. However, the X-ray details embedded within the bright and dark regions cannot be enhanced at the same time in practically. Therefore, in this paper, we make an assumption that an X-ray image could be decomposed into tissue components and important details. The extraction of tissue components was then posted as a contrast maximization problem. Next, we aimed to adaptively adjust the attenuation ratios of tissue components over the image. Since tissues may not be the major focus of X-ray inspection, we proposed to enhance the visual contrast of detail components by tissue attenuation and dynamic range stretching. To realize our system, a parametric adjustment model was deduced to efficiently generate many enhanced images in a global fashion. Later on, an ensemble framework was finally proposed to fuse multiple global enhanced images and produce the vivid output by maximizing the visual contrast in both bright and dark regions. We have used four measurement metrics to evaluate our system and reach promising scores. Moreover, we applied our system to an X-ray dataset provided by Japanese Society of Radiological Technology. The experiment results also demonstrate the effectiveness of our method in X-ray image enhancement.

Demo
Input Image	Output Image

*We've provided a testing dataset here, please feel free to use it if you need.


X-Ray input image:


Intput image	Output image

The figure below shows the summary of score from 12 valid questionaires of each method .

The following figures represent the score distributions of method corresponding to their title, and The blue color part of bar shows the score distribution in first group, red one represent for the score distribution second group, and so on.