Search

Object Stereo Vision has conventionally been one of the deeply examined areas in computer vision. Stereo matching is employed in numerous modern applications, including robot navigation, augmented reality, and automotive applications. Even though it has a long research history, it is still challenging for the edges of textureless, discontinues, and occluded regions under radiometric variation. This research article proposes a modified histogram equalization, a novel feature extraction, a spatial gradient model, and matching cost, which is robust and stable to images taken in different radiometric variations. The proposed method reduced the average percentage of bad pixels to 3.35 and reduced the relative mean square error (RMSE) up to 30.08 on the Middlebury dataset for different illumination and exposure values. Quantitative and qualitative evaluation of the proposed method demonstrates significant improvement in increasing PSNR and decreasing bad pixel percentage against radiometric variation and state-of-the-art local stereo matching algorithms.

As the techniques of autonomous driving become increasingly valued and universal, real-time semantic segmentation has become very popular and challenging in the field of deep learning and computer vision in recent years. However, in order to apply the deep learning model to edge devices accompanying sensors on vehicles, we need to design a structure that has the best trade-off between accuracy and inference time. In previous works, several methods sacrificed accuracy to obtain a faster inference time, while others aimed to find the best accuracy under the condition of real time. Nevertheless, the accuracies of previous real-time semantic segmentation methods still have a large gap compared to general semantic segmentation methods. As a result, we propose a network architecture based on a dual encoder and a self-attention mechanism. Compared with preceding works, we achieved a 78.6% mIoU with a speed of 39.4 FPS with a 1024 × 2048 resolution on a Cityscapes test submission.

In order to improve the accuracy of rotated text detection effectively, we combine the network architectures with different functions to make up for the shortcomings of each network. First, rotate proposal regions to achieve scene text detection in any direction. We also enhance the features of important channels and weaken the features of non-important channels to effectively improve the accuracy of the overall network. In order to achieve multi-scale target detection in the scene, the chips of the same size are intercepted on different scales. Therefore, it can zoom in on smaller targets, and zoom out on larger targets. The experiments show that this method is competitive in accuracy with traditional architecture.

This paper reviews the NTIRE 2020 Challenge on NonHomogeneous Dehazing of images (restoration of rich details in hazy image). We focus on the proposed solutions and their results evaluated on NH-Haze, a novel dataset consisting of 55 pairs of real haze free and nonhomogeneous hazy images recorded outdoor. NH-Haze is the first realistic nonhomogeneous haze dataset that provides ground truth images. The nonhomogeneous haze has been produced using a professional haze generator that imitates the real conditions of haze scenes. 168 participants registered in the challenge and 27 teams competed in the final testing phase. The proposed solutions gauge the state-of-the-art in image dehazing., CVPR Workshops Proceedings 2020

This work presented an end-to-end training model for parking slot detection in automatic parking systems (APSs), which combined both a line and a point semantic segmentation models based on multi-task learning. The proposed models generate images of entrance line and center points of corners, and are used to determine the coordinates of actual parking slots in the post processing step. The recall, precision and F-measure rate of the proposed method are 92.94%, 99.40% and 96.06%, respectively, which are better than existing state-of-the-art methods with end-to-end training.

In video surveillance, person re-identification is an important task of recognizing individuals in diverse locations over different non-overlapping camera views under the condition of large illumination variations. In order to deal with these challenges, two different and efficient color-based-methods are proposed for single-shot person re-identification in this article, which uses the Gaussian mixture model to combine with color histograms (low-level feature) and the dense salient patches (mid-level feature) as the color features. Both the two proposed systems are three-stage processes. The first stage is the image enhancement by illumination normalization, and it used to deal the intensity variations. The second stage includes pedestrian segmentation and human region partition, which separates the background (BG) and foreground (FG) and locates the body segments to improve the accuracy of feature extracting and matching. The third stage is to perform feature extracting and matching. A Gaussian mixture model is used in the first system, GMMWCH, to generate the weighted color histogram as the color features, which has a low computation time and a good recognition rate. For the second system, SaliGMMWCH, the dense correspondence is used to link the color histogram weighted by the Gaussian mixture model to find salient regions. Even though that takes more time for computation, the SaliGMMWCH retains a better recognition rate than GMMWCH. In addition, the correct match can be chosen by matching the similarity scores of different feature with an appropriate weight selection. Both the proposed methods have been tested on the benchmark, VIPeR and PRID 2011, for evaluation. The experimental results demonstrate superior recognition rate and execution performance by using the proposed methods compared to other representative methods.

Human-robot interaction has recently drawn great attention in both academic and industrial communities. Given the benefit features of human-robot interaction, this article developed a robot navigation control system with graphic user interface using user’s hand positions and hand gesture recognition. A new skin color fusion of channels among several color spaces, such as RGB-H-Cr-Cb, is proposed in this research. Human face properties are taken as a reference to detect and segment the hand regions. The hand recognition module uses the Weber local descriptor cooperated with the support vector machine. In experimental results, the skin segmentation as well as hand recognition method is compared with some previous methods to demonstrate the proposed methods’ flexibility and robustness to be implemented in real-time applications.

In order to study and observe images, image enhancement methods are often used to improve low quality images. The purpose of image enhancement is to make the region of interest become clearer, and to weaken or even delete the uninterested regions. In this paper, we use the retinex theory and exposure fusion to enhance the contrast of images. In the proposed method, we only manipulate the V channel of the HSV color space to maintain color consistency as far as possible. First, we estimate the local and the global incident lights respectively, and use the retinex theory to get images with local and global details. Following exposure fusion, the two images and the original image are used to generate the enhanced image. Experiments show that the proposed method can strengthen both the local and global details, and effectively reduce halo, color distortion, and noise problems. In short, the result images look quite nature. In summary, the proposed method can enhance the image quality, and make images easier to be analyzed and understood.

In this article, a dynamic gesture recognition system with the depth information is proposed. The proposed system consists of three main components: preprocessing, static posture recognition and dynamic gesture recognition. In the first component, the background subtraction is used to exclude invalid gestures that is not generated by the main user, and then to detect and track the hand. Second, the region of hand is extracted using an adaptive square. Once the region of hand is obtained, the features of hand are extracted and the static hand posture are classified using the support vector machine (SVM). Finally, nine commonly used dynamic hand gestures can be detected using different methods. In the experiments, the static hand posture classification was evaluated in different postures and the performance of dynamic gesture recognition is verified by two different persons at 4 different position with 2 different depths. The experiment results show that the proposed system can accurately detect the dynamic hand gestures with an average recognition rate of 87.6 %, which is good for controlling the embedded systems, such as home appliances.

In this paper, we proposed a seam carving based refinement method to refine and produce superpixels. The proposed method can refine existing superpixels by repeating the splitting process. There are two major steps. The first is choosing a superpixel candidate by analyzing color variances; the second is splitting a superpixel into 4 by dynamic programming. The experimental results show that the proposed method can achieve higher accuracy compare to original methods in the same number of superpixels.

Lane-mark detection is one of the most important parts of the intelligent transportation system (ITS). Cameras mounted in front of automobiles are used to capture the road scene to detect lane-marks. This paper proposes a new lane-mark extraction algorithm that consists of six parts. We first determine the regions of interest (ROIs) of captured images and investigate their boundaries. Then, we divide the boundary image into subimages to calculate the local edge-orientation of each block and remove edges with abnormal orientations. After that, the multi-adaptive thresholding method is employed for each subimage so that the system can work efficiently for an image that has different contrasts to indicate sub-regions for lane-marks. We then verify these candidate lane-mark edges and fit the lane-marks using either straight or curved line models. A Kalman filter is also applied to track them. The proposed method is evaluated in several situations, including in bad weather conditions and in the presence of shadow effects, obstacles such as wipers, or road signs on the road. The results show that the proposed method can detect the lane-marks in real time for various complex environments.

This article describes a novel scheme to save the usage of ink or toner by an approximate-K algorithm for color printers. Existing printers use the mixtures of three color toners (Cyan, Magenta and Yellow) to print all the pixels for color images, and it makes color printing 4---4.5 times more expensive than monochromic printing. Since human eyes are not sensitive to distinguish neighboring colors in the color space, we can use the K (blackK) toner to replace the colors close to gray-scale. We can then reduce the ink usage without affecting the image visual qualities. We use the saturation in the HSV (hue, saturation, value) color model to discover the near gray-scale pixels and transform those pixels to gray level. We then evaluate the objective image quality using the PSNR (Peak Signal Noise Ratio) and use the DSCQS (Double Stimulus Continuous Quality Scale) as the subjective evaluation method. From our experimental results, printing a color image using our algorithm needs only 84 % of the original price in average.

A new method for detecting text in video images is proposed in this article. Variations in background complexity, font size and color, make detecting text regions in video images a difficult task. A pyramidal scheme is utilized to solve these problems. First, two downsized images are generated by bilinear interpolation from the original image. Then, the gradient difference of each pixel is calculated for three differently sized images, including the original one. Next, three K-means clustering procedures are applied to separate all the pixels of the three gradient difference images into two clusters: text and non-text, separately. The K-means clustering results are then combined to form the text regions. Thereafter, projection profile analysis is applied to the Sobel edge map of each text region to determine the boundaries of candidate text regions. Finally, we identify text candidates through two verification phases. In the first verification phase, we verify the geometrical properties and texture of each text candidate. In the second verification phase, statistical characteristics of the text candidate are computed using a discrete wavelet transform, and then the principal component analysis is further used to reduce the number of dimensions of these features. Next, the optimal decision function of the support vector machine, obtained by sequential minimal optimization, is applied to determine whether the text candidates contain texts or not.

In this paper, we proposed a dynamic hand gesture recognition system by using only the depth information. The proposed system can recognize twelve different dynamic hand gestures, including swipes, scales, push, wave, rotates, circle, and drag. First, the background subtraction is used to remove the unnecessary information, and the depth information of main user can be obtained. Furthermore, hand position can be tracked, and the region of hand is extracted as an adaptive square. Once the region of hand is obtained, the hand parameters are obtained by calculating the depth information of hand region. The proposed system can recognize dynamic hand gesture by using the hand parameters. In the experiment, the performance of the proposed system is verified by two different people at 2 different depths, and both right and left hands are verified. The experimental result show that the proposed system can recognize the dynamic hand gestures with an average recognition rate of 90.08%.

Embedded device recently becomes more crucial in some real-life applications, such as the entertainment, health care, and industrial application. Given the development of embedded devices, the mobile robot becomes not only cheaper, but also more highly practical to produce. However, the embedded device still has some limitation from the hardware capability, therefore the implemented method has to be robust and low computational complexity. In this research, we introduce an embedded human following mobile robot application with an RGB-D camera, which can help us detect human position based on the human geometry, compatible into the ARM-based embedded device. The Human Depth of Interest (HDOI) is proposed to cooperate with CAM-shift algorithm for the human detection and tracking. The virtual spring model with the safe region and active region is chosen to carry out the smoothing navigation and easy implementation. The performance of our approach is enhanced more accuracy than the original CAM-shift algorithm and more feasible than the stereo vision system.

The objective of the proposed method is to reconstruct the three-dimensional structure of a scene in real-time via moving a low-cost RGB mono-camera by hand. Only simple and fast algorithms are used for the purpose of real-time with a decent structure resolution. To compensate the defects of the cheap algorithms, the scene structures will be buffered and used only when they are considered as a good case, and they will be smoothed to decrease the effects of the structural defect. Even if only a minority of overall cases are good cases, a complete scene structure can still be obtained with small amounts of different views. The proposed method can be applied into embedded systems after optimization and adjustment.

Pedestrian detection is an important research field in advanced driver assistance system (ADAS). This paper puts forward a pedestrian detection framework based on both heuristic statistics and machine learning. First, a restriction of region of interest (ROI) is set on the captured image. Second, the template matching coarsely detects candidate pedestrians by using a set of template images, the edge image of the current frame, and the difference image from previous and current frames. Next, the histogram analysis again roughly filters out the candidate pedestrians. Finally, Histogram of Oriented Gradients (HOG) combined with library support vector machine (LIBSVM) is used to verify those candidate pedestrians. The experimental results show that the proposed method can run in real-time, where the false negative rate is 1.43%, and the false positive rate is 0.16%.

In video surveillance, person re-identification is an important task of recognizing individuals in diverse locations over different non-overlapping camera views under the condition of large illumination variations. To deal with these challenges, an efficient appearance-based-method was proposed for the single-shot person re-identification, that use a mixture of Gaussian models to weight HSV color histograms as color features. The proposed approach has been tested on a public benchmark dataset, VIPeR, for evaluation. The experimental results demonstrate superior recognition rate and execution performance by using the proposed method compared to the other representative methods.

We propose a method to generate a wider Field of View (FOV) using multiple RGB-D cameras. In the initialization step, we extract the matched feature points from the moving object in order to calculate the transformation matrix. Then, the depth data would be aligned using the transformation matrix in the world coordinate system. The wider FOV is generated after the perspective projection. The results show that two input depth images can be correctly combined into a wider depth image.

This paper presented a dynamic hand gesture recognition system for home appliance control using only the depth camera. The dynamic hand gesture is recognized using static hand postures and hand trajectory. The proposed system can recognize seven commonly used dynamic hand gestures. Experimental results show that the system is effective for home appliance control.

Smoking in public indoor spaces has become prohibited in many countries since it not only affects the health of the people around you, but also increases the risk of fire outbreaks. This paper proposes a novel scheme to automatically detect and recognize smoking events by using exsiting surveillance cameras. The main idea of our proposed method is to detect human smoking events by recognizing their actions. In this scheme, the human pose estimation is introduced to analyze human actions from their poses. The human pose estimation method segments head and both hands from human body parts by using a skin color detection method. However, the skin color methods may fail in insufficient light conditions. Therefore, the lighting compensation is applied to help the skin color detection method become more accurate. Due to the human body parts may be covered by shadows, which may cause the human pose estimation to fail, the Kalman filter is applied to track the missed body parts. After that, we evaluate the probability features of hands approaching the head. The support vector machine (SVM) is applied to learn and recognize the smoking events by the probability features. To analysis the performance of proposed method, the datasets established in the survillance camera view under indoor enviroment are tested. The experimental results show the effectiveness of our proposed method with accuracy rate of 83.33%.

Lane-mark detection is one of the most important parts in intelligent transportation systems (ITS). In this paper, we propose a lane-mark detection system which can overcome a lot of difficult situations, such as bad weather conditions, shadow effect, or road sign on the road. After the region of interest (ROI) of a road image is determined, we apply the Canny edge detector to investigate boundaries. In order to remove the noise edges, we divide the boundary image into sub-images to calculate local edge-orientation of each block and remove the edge with abnormal orientation. In this step, we produce a table to store the blocks which satisfy the assumption of lane-mark edge-orientation and use this information as an adaptive ROI of lane-marks. We propose the edge-pair scanning method to verify the edges which belong to lane-marks by using the relationship of adjacent edges of lane-marks and the width between these two edges. In the local adaptive threshold finding method, we also divide the image into sub-images and apply the feature that road lane-marks are always painted with high contrast colors with the road surface. Then, we use multi-adaptive thresholding method for each block. The system can work robustly under the situation that different parts of the image have different contrast for lane-marks. After eliminate the noise edges, we apply Hough Transform to fit the lane-marks as straight line models. The experiment results show that the proposed method can detect the lane-marks in real-time for various different environments.

This paper proposes a novel scheme to save the usage of ink or toner by an approximate-K algorithm for color printers. Existing printers use the mixtures of three color toners(C, M and Y) to print all the pixels for color images, and it makes color printing cost 4-4.5 times more expensive than monochromic printing. Since human eyes are not sensitive to distinguish neighboring colors in the color space, we can use the K toner to replace the colors close to gray-scale. We can then reduce the ink usage without affecting the image visual qualities. We use the saturation in the HSV color model to discover the near gray-scale pixels and transform those pixels to gray level. We then evaluate the objective image quality using the PSNR. From our experimental results, printing a color image using our algorithm needs only 77% of the original price in average.