Recent advances in artificial intelligence and robotics have significantly expanded the capabilities of autonomous navigation systems in the industrial and service applications. However, most existing navigation pipelines mainly rely on accurate metric maps and specialized sensors, such as LiDAR or depth cameras, which constrain scalability in dynamic or previously unseen environments and incur substantial deployment and maintenance costs. This paper presents a zero-shot object goal navigation framework, STG-Nav, which leverages large language models (LLMs) and topological graph representations to enable semantic navigation without pre-established maps or explicit goal annotations. The proposed system integrates online semantic mapping, skeletal topological graph (STG) construction, grounded language-image pretraining (GLIP), and LLM-driven semantic reasoning for node evaluation and selection. Using visual observations acquired during exploration, the robot incrementally constructs a topological graph, infers node semantics, and performs object-conditioned semantic scoring to guide navigation decisions. By combining multi-view visual perception with language-grounded reasoning, the framework enables semantic filtering and goal inference in unknown environments, and provides robust object-centric navigation under zero-shot conditions. Compared to conventional map-based and sensor-fusion approaches, the proposed method significantly reduces sensor dependence and system complexity while improving the adaptability to complex and cluttered scenes. Experimental results demonstrate stable exploration behavior and superior target-guided navigation performance across diverse environments, validating the effectiveness of the proposed approach for zero-shot object goal navigation. Code is available at https://github.com/ting0606/STG-Nav .

The segmentation of road markings plays a crucial role in visual perception for the autonomous driving system. It enables vehicles to recognize road markings at the pixel-level, and facilitates subsequent path planning, localization, and map construction tasks. Current techniques mainly focus on normal driving scenes (i.e., clear daytime), and the performance would decrease significantly for adverse weather conditions. This work proposes RMSeg-UDA: an unsupervised domain adaptive road marking segmentation framework. By combining schedule self-training and class-conditioned adversarial training, the network utilizes both labeled normal data and unlabeled data from other domains to train a road marking segmentation model. For the evaluation on adverse conditions, a new image dataset, RLMDAC, is established with rainy and nighttime driving scenes. The experiments conducted using both public and our datasets have demonstrated the effectiveness of the proposed technique. Code and dataset are available at https://github.com/stu9113611/RMSeg-UDA.

Traffic sign detection plays an essential role in advanced driver assistance system (ADAS) or self-driving vehicles. Typically, deep neural networks are employed to analyze road scene images captured by an onboard camera. However, due to the significant variation in appearance of different traffic signs, the classification of high similarity patterns is still a challenging task. To address these issues, this paper presents an end-to-end traffic sign detection framework based on DETR. The proposed network incorporates data augmentation and negative sample learning to mitigate the problem of data imbalance and enhance the model recognition capability effectively. An UASPP module (Upsample Atrous Pyramid Pooling) is introduced to integrate multi-scale features and global information. In the experiments, the performance evaluation has demonstrated the improvement of mAP by 3.9% on TT100K and 36.3% on GTSDB compared to state-of-the-art methods. The code and datasets are available at https://github.com/chinglun/TS-DETR.

With advancement of robotic technologies, research and developments of humanoid robots have become increasingly widespread. During task execution, it generally requires a SLAM system for robot self-localization and environment map construction. However, humanoid robots face the challenges of image blur and erroneous feature matching caused by the irregular bipedal motion and gait trajectories. This paper presents HuKIS-SLAM, an image stabilization neural network model which incorporates the kinematics of humanoid robots for simultaneous localization and mapping. It takes the humanoid robot’s head pose and shaky video stream as input to generate stabilized image sequences. The output frames are employed for consistent 3D reconstruction and enhancing the accuracy and robustness of robot localization. We evaluate the proposed method and state-of-the-art SLAM systems on public TUM and our humanoid robot datasets. Experimental results demonstrate that our technique performs stable tracking even in shaky image sequences, and achieves minimal precision errors in camera trajectory The code and datasets are available at https://github.com/KangKaiChen/HuKIS.

The rapid evolution of advanced driver assistance systems (ADAS) has been driven by the advances of deep neural networks, and multi-tasking is essential for autonomous driving systems. This paper presents FuseRoad, a new multi-task model that leverages cross-dataset learning to address the dependency on specific multi-task datasets and reduce the annotation costs. It integrates semantic segmentation and lane detection into an end-to-end framework while providing an effective approach to utilize multiple single-task datasets. By incorporating Semantic Road Knowledge Extractor (SRKE) to direct more attentions on the roadway, FuseRoad enhances the accuracy and reliability of lane detection. The model also employs the logit normalization loss to address the issue of overconfidence commonly faced by conventional lane detection methods. In experiments, FuseRoad outperforms state-of-the-art approaches in both accuracy and F-1 score. The evaluation on semantic segmentation metrics also demonstrates that the proposed technique is highly effective for multi-task road scene analysis. Code and datasets are available at https://github.com/HengChihHsiao/FuseRoad.

With the advances of deep neural networks, there is progress on the detection and recognition of traffic lights for advanced driver assistance systems (ADAS). However, existing approaches most rely on the identification of traffic light boxes, followed by the recognition of signal lights. It is considered as a major drawback since light bulbs can be arranged in different directions or irregular patterns in different geographic regions. In this paper, we present a traffic light detection method based on direct recognition of individual signal lights. Our two-stage technique utilizes data augmentation and ensemble learning to detect the light bulbs with least miss rate. By learning the color characteristics from validation sets for data augmentation, it is able to achieve a signal light candidate detection rate at 97.26%. Followed by the classification stage, the recognition accuracy is given by 98.6%, which outperforms state-of-the-art traffic light detection algorithms. The source code and dataset are available at https://github.com/981124/yolov7 traffic light detect.

With the development in sensors, computing resources and deep neural networks, the safety mechanisms of vehicles are continuously developed towards the fully automated driving system (FADS). One of the most crucial aspects of this technology is the environmental perception. Most of the existing works focus on recognizing specific targets in the scene, and often overlook the holistic information for sufficient use by FADS. In this paper, we adopt a multi-task learning approach to achieve more comprehensive recognition of environmental information. On the other hand, due to the increasing prominence of traffic issues in urban areas, the problem of illegal parking has gained more attention. Thus, a vision system to recognize illegal parking in road scenes based on environmental perception is proposed. We also collect an illegal parking dataset and make it available publicly for related research. Source code and dataset are available publicly.

With the continuous advances of technologies, the demand for precise vehicle positioning has grown significantly. Although it is now possible to capture driving scenes and record driving paths using a dashcam, standard civilian GPS typically has the accuracy errors ranging from 3 to 5 meters. This level of precision is in general not sufficient for the rapidly evolving ADAS (Advanced Driver Assistance Systems). In this paper, we present a high-precision vehicle positioning technique based on satellite map and image data. Instead of using expensive LiDAR sensors, the proposed approach utilizes lane detection, semantic segmentation and geolocation to extract environmental features from images. A* algorithm is then adopted to refine the driving trajectory for the improvement of vehicle positioning accuracy. Furthermore, we establish an image dataset containing satellite maps and latitude/longitude coordinate information of various road scenes. The code and datasets are made available publicly at https://github.com/M6104150181M610415018-Paper.

With the advancement of computational intelligence, autonomous driving has become the future development trend of the automotive industry. Since the safety is commonly considered as the first priority of self-driving and driver assistance systems, the understanding of transportation infrastructure is an essential problem. In this paper, a technique for traffic sign detection and recognition is proposed. Different from the general image-based methods, we also incorporate the vehicle position and geographic information to improve the accuracy. Based on the approximate location of a traffic sign obtained from GPS and GIS, it can be used to increase the confidence level of network detection results. In the experiments, the training datasets are derived from Google Street View images and collected with an in-vehicle camera. The performance evaluation compared to the image-only methods has demonstrated the effectiveness of the proposed approach.

Driver distraction detection is an important function of driver monitoring systems and intelligent vehicles. Most previous research only focuses on the system development for daytime operations. In this paper, we propose a network model, V2IA-Net, which is able to use the daytime visible and nighttime infrared images for the driver distraction detection task. With the visible-infrared image translation, driver action recognition and head pose detection, the driver distraction behavior can be analyzed in real-time performance. To provide realistic driving scenes for network training and testing, a visible-infrared image dataset, VID, is created. The proposed V2IA-Net is trained on the unpaired images, and capable of common feature extraction for visible-infrared image conversion. In the experiments, our technique is compared with various driver distraction detection models. The results have demonstrated the effectiveness of the proposed method. Source code and datasets are available at https://github.com/kk2487/V2IA-Net.

Different applications for advanced driver assistance systems (ADAS) have been increasing rapidly. With the advances of machine learning techniques, traffic risk assessment is becoming possible by understanding the complex driving scenes. This paper presents a technique to classify possible risks in four levels using deep neural networks. The important regions in road scene images are first predicted, followed by the risk assessment using a series of weighted RGB and optical flow images obtained from the proposed network model. To achieve better understanding of driving scenes, driver visual attention is incorporated to enhance the feature extraction. An LSTM model is then employed to learn the temporal information from road scene video clips. In addition to the experiments using the public TRA dataset, we also present a new LTDR dataset for performance evaluation. Compared with existing techniques, the result has demonstrated the effectiveness of the proposed method. The code is available at https://github.com/steven-wj-wu/RVL_Traffic_Risk_Assessment.

This paper presents a method to improve the localization accuracy for visual inertial odometry (VIO) by combining the ultra-wideband (UWB) positioning technology. The overall architecture is mainly divided into two stages. In the first stage, the constraint on UWB short-term position change is adopted to improve the pose estimation results of the VIO system. It is also used to solve the translation error caused by the lack of visual features and vibration during the flight. In the second stage, a loose coupling method based on nonlinear optimization is utilized to fuse the local pose estimator of the VIO system with the global constraints from the UWB positioning. At the beginning of each optimization, the alignment between the VIO and UWB frames is estimated to avoid the influence of the coordinate transformation caused by the cumulative error of the VIO system. Since there are no public datasets available for comparison, we have established several datasets containing large vibration amplitudes and weak feature points. In the experiment, the performance evaluation of the proposed technique is carried out using the Apriltag approach for verification.

Under the recent trend of unmanned economy, the retail stores have reduced the manpower for service and cashier gradually. The retail product recognition becomes one essential problem for unmanned shopping. Although the success of deep neural network makes the object recognition feasible in various applications, it is still difficult to perform well on a large number of classes. This paper presents an ensemble learning approach to deal with recognition for the growing number of retail products. In the proposed technique, the object classification networks are first improved with feature extraction and block attention. The ensemble model is then constructed by integrating the multiple network models with the loss selection as model weights. In the experiments, the feasibility of our ensemble recognition method is validated with a number of production items. The results have demonstrated the effectiveness compared to the state-of-the-art recognition algorithms.

Due to the recent progress on autonomous driving, some technologies have been gradually deployed to the public transport vehicles. The passenger safety under the unmanned operating environment has become an emerging issue which requires much more attention. This paper presents a method for passenger detection, counting and action recognition inside a minibus. A top-view camera system is mounted on the ceiling to have a full coverage of the interior. 2D and 3D convolutional neural networks are developed for pose recognition and action classification of the passengers. The experiments are carried out in a self-driving minibus, and the results have demonstrated the feasibility of the proposed technique.

Understanding 3D scenes have attracted significant interests in recent years. Specifically, it is used with visual sensors to provide the information for a robotic manipulator to interact with the target object. Thus, 6D pose estimation and object recognition from point clouds or RGB-D images are important tasks for visual servoing. In this paper, we propose a learning based approach to perform 6D pose estimation for robotic manipulation using Mask R-CNN and the structured light technique. The proposed technique optimizes the 6D pose between the target objects and 3D CAD models in multi-layers. Our method is evaluated on a publicly available dataset for 6D pose estimation and shows its efficiency in computation time. The experimental results demonstrate the feasibility of the random bin picking application.

Semantic segmentation and shape detection of lumbar vertebrae, sacrum, and femoral heads from clinical X-ray images are important and challenging tasks. In this paper, we propose a new multi-path convolutional neural network, BiLuNet, for semantic segmentation on X-ray images. The network is capable of medical image segmentation with very limited training data. With the shape fitting of the bones, we can identify the location of the target regions very accurately for lumbar vertebra inspection. We collected our dataset and annotated by doctors for model training and performance evaluation. Compared to the state-of-the-art methods, the proposed technique provides better mIoUs and higher success rates with the same training data. The experimental results have demonstrated the feasibility of our network to perform semantic segmentation for lumbar vertebrae, sacrum, and femoral heads. Code is available at: https://github.com/LuanTran07/BiLUnet-Lumbar-Spine.

The visual perception is of great significance for advanced driving assistance systems or autonomous driving vehicles to recognize the surrounding scenes. In the adaptation to the real environments for collision warnings, a sensor system should be efficient and has the strong ability to detect small objects. This paper presents a forward collision warning technique which incorporates the object detection and depth estimation networks. A deep convolutional neural network is constructed with transfer connection blocks for object detection and classification. It is capable of small object detection under the real-time processing requirement. For depth estimation, a monocular based disparity estimation network is adopted to the stereo vision framework. The epipolar constraint is applied to increase the prediction accuracy. In the experiments, the performance evaluation is carried out on public driving datasets. The comparison with the state-of-the-art networks has demonstrated the feasibility of the proposed technique.

The driving time and usage of reverse gear are much less than other driving situations, but the dangers are easy to be ignored due to of the blind spots when driving the car in reverse gear. The safety in the rear is always an important issue for vehicles. In this work, a rear obstacle detection technique is proposed to remind the drivers when moving backwards. We use a binocular camera system for rear obstacle detection based on parallax map. An image representation called stixel is adopted for the outdoor traffic scene in a columnar manner. The disparity map is used to estimate the ground region, identify the free space, and segment the obstacles based on the height. Finally, a road edge tracking step is incorporated to to stabilize the system. In the experiments, several reverse parking video sequences are acquired to validate the effectiveness of the proposed algorithm.

With the rapid development of advanced driving assistance technologies, from the very beginning of parking assistance, lane departure warning, forward collision warning, to active distance control cruise, the active safety protection of vehicles has gained the popularity in recent years. However, there are several important issues in the image based forward collision warning systems. If the characteristics of vehicles are defined manually for detection, we need to consider various conditions to set the threshold to fit a variety of the environment change. Although the state-of-art machine learning methods can provide more accurate results then ever, the required computation cost is far much higher. In order to find a balance between these two approaches, we present a detection-tracking technique for forward collision warning. The motion tracking algorithm is built on top of the convolutional neural networks for vehicle detection. For all processed image frames, the ratio between detection and tracking is well adjusted to achieve a good performance with an accuracy/computation trade-off. Th experiments with real-time results are presented with a GPU computing platform.

Due to the recent development of industrial automation, using vision based techniques to estimate the pose of workpieces for random bin picking application is a future trend. The common method for 3D pose estimation is to use the CAD model and feature points. However, the CAD-based method has a high degree of flexibility in the assembly line, and it is not easy to acquire all necessary features of the workpieces. In this work, we use two depth cameras to capture the 3D scene, and propose a CAD-based multi-view pose estimation algorithm. First, RANSAC and an outlier filter are adopted for noise removal and object segmentation. We use a voting scheme for preliminary pose estimation, followed by the ICP algorithm to derive a more precise target pose. Finally, with disturbance detection, the robot arm can grip the objects without rescanning for each operation. A complete system for 3D scene acquisition using structured light cameras, 3D pose estimation and robot arm control is developed for the pick-and-place task. Experiments are carried out in the real scene environment to demonstrate the feasibility of the proposed technique.

This paper investigates stereo matching with zooming information. In the proposed technique, two zoom lens cameras are used to capture multiple stereo image pairs with different focal length. With the assistance of zoom image pairs captured from the same camera, it is possible to increase the correctness of stereo matching results. In addition, a zoom rectification method is proposed to simplify the zoom vector computation. In the experiments, we use the bad pixel rate to compare the results obtained from the real scene images using conventional stereo matching techniques and the proposed stereo with zooming approach. Moreover, we test the proposed algorithm on the real image with optical zoom to demonstrate the effectiveness of our work. It is shown that our technique is able to reduce the bad pixel rate using both the local and global stereo matching algorithms.

With the revolution of Industry 4.0, the development of Cyber-Physical Systems (CPS) has become a major research topic. The objective of this paper is to develop an intelligent automatic operational management platform. This platform will be used to collect data, provide predictive maintenance, and monitor and control the robots. The performance evaluation strategy is used to improve the efficiency of the management system. The experiments with battery simulation warning and manufacturing execution are presented to demonstrate the rudimentary system design.

This paper presents an early research on the route recommendation system based on street view images. It consists of three primary functions. (1) Data collection: The user selects a path using a graphical user interface and download all street view images for the path. (2) Road surface extraction: The vanishing point in the image is obtained according to the road surface region segmentation by local growing at a super-pixel level. (3) Road surface identification: The images of brick and asphalt road surfaces are collected for training, and then used to identify the road surface type for route recommendation. The objective is to find the most comfortable path, which is not necessarily to be the shortest one. The experimental results show that our pre-selected groups of route well suit the original expectation.

This paper presents a new stereo configuration based on sensor rotation. Different from the traditional stereo vision system which consists of multiple cameras, our rotational stereo device is built with a single image sensor. By constraining the sensor's planar motion perpendicular to the optical axis with an offset to the axis of rotation, the visual parallax can be generated with multiple viewpoints. The stereo image pairs captured with sensor rotation are used for correspondence matching and disparity computation. A variation of multiple baseline stereo with radial configuration is proposed. It is able to provide more flexible stereo baseline settings and improve the accuracy of depth measurement. The geometric formulation and computational algorithms are developed, and a prototype range sensing device is constructed. Experimental results obtained from the real scene environment have demonstrated the feasibility of the proposed technique.

In orthopedic surgery, three-dimensional models are usually reconstructed through CT or MRI images. In urgent operations, the most common way for 3-D model reconstruction is to extract C-arm images by C-arm fluoroscopic imaging. However, it is harmful to human bodies because of the ionizing radiation. In this work, we present a model deformation approach for 3D reconstruction from medical imaging. Our approach utilizes the visual hull algorithm to reconstruct a rough 3-D model from C-arm images, and then deforms it according a reference model. During the deformation stage, some adjustments are made for the models and the coherent point drift algorithm is carried out to match the 3-D points of two models. Experimental results demonstrate the feasibility of our technique for 3-D model reconstruction from C-arm imaging.

This paper presents a new stereo matching technique based on image bit-plane slicing and fusion. Given an 8-bit image, it can be separated to multiple images with 1-bit representation by bit-plane slicing. Instead of performing stereo matching on the original image pair, bit-plane slices are used to find stereo correspondences, followed by image fusion for the final disparity map. The main advantage in our method is the extremely low data access requirement, which is an important issue for high dynamic range (HDR) image applications or stereo matching implementation on embedded systems. The proposed method can adapt to the existing stereo matching techniques such as local block matching (SAD), semi-global block matching, probabilistic correspondence matching (SSMP), and global optimization (belief propagation). In the fusion stage, we investigate wavelet (in the transform domain) and PCA (in the spatial domain) methods. Experiments are carried out on Middlebury datasets for different stereo matching techniques with full bit-rate matching. The results demonstrate that the bad pixel rate can be reduced by our bit-plane matching and disparity fusion framework.

Scene recognition and localization are important research topics for driver assistance technology and autonomous mobile robot in recent years. In this paper, we present a novel system which is able to detect the node information and construct the topological map based on . We use the node information from the topological map for image retrieval and localization. For the topological map construction, we utilize the Extended-HCT method and feature extraction. We also combine the content-based and feature-based image retrieval techniques for recognition and localization in the real scene image dataset. By using the proposed approach, we are able to construct a real-time image retrieval system for navigation assistance, and verify the correctness of the route. The experiments carried out on our dataset demonstrate that the proposed approach improves the conventional retrieval methods.

In recent years, many works have addressed the issues of generating high dynamic range (HDR) images from the low dynamic range (LDR) counterparts. Since the HDR image contains a broader range of physical values which cannot be recorded by conventional sensors, the previous approaches use a sequence of images captured with different exposures to synthesize an HDR image. In this paper, we propose a spatial down-sampling technique to extend the dynamic range of an LDR image and generate an image with a broader dynamic range. The idea is to trade the large resolution of an image with a large brightness range for intensity quantization, and produce a so-called Extended Dynamic Range (EDR) image. Experimental results demonstrate that our approach is able to provide the better image quality than those derived from the existing LDR to HDR image conversion techniques.

This paper presents geometric techniques for self-localization improvement, especially for the robots equipped with a single catadioptric camera. We take the vertical line and intersection point matching into account, and proposed a novel descriptor named “Double-Gaussian vector”. The vector uses two Gaussian matrices to blur the process image region and build the corresponding feature vectors for solving the vertical line matching in two consecutive video frames. For ground plane estimation, the perpendicular lines with respect to optical axis are extracted by two approximate curve equations. The equations then crop the ground plane area of the omnidirectional image. The sparse bundle adjustment (SBA) is adopted for iterative calculating the 3D matching points between two robot locations for optimizing the robot pose estimation. The convergent 3D points are used to compute the robot poses and record the navigation trajectory. The results show that the proposed methods significantly improve the robot localization and navigation compared to the previous literature in the experiments.

Free-viewpoint video is to enable the viewer to choose arbitrary viewpoints for the scene captured by a multi-view imaging system. In this paper we present a multi-layered variate-resolution sampling technique for 3D scene representation. The 3D point cloud obtained from the reconstructed 3D scene is used for novel view rendering. For any given viewpoint, image synthesis with different level of detail is carried out using the quadtree based non-uniformly sampled 3D data points. Experimental results are presented using the 3D model of a reconstructed real object.

This paper presents a novel encoding method for scene change detection and appearance-based topological localization framework. The relation computation over convex hull points is used to compare the similarity between the scenes. It relies on the relative ordering of the feature strength, not directly on the feature vectors. We first deal with multiple convex hulls over the detected features and then compile statistics for coding on the hull points through a vector magnitude comparison. Finally, the hull points are formed by binary codes. The codes are suitable for scene change detection and visual place recognition by statistical analysis. The experimental results show the coding method is robust under the varying environment.

This paper presents a method for generating high dynamic range and disparity images by simultaneously capturing the high and low exposure images using a pair of cameras. The proposed stereoscopic high dynamic range imaging technique is able to record multiple exposures without any time delay, and thus suitable for high dynamic range video synthesis. We have demonstrated that it is possible to construct the camera response function using a pair of images with different amount of exposure. The intensities of the stereo images can then be normalized for correspondence matching. Experiments using the Middlebury stereo datasets are presented.

A human pose estimation method from monocular image captures is presented. The objective is to develop a human-computer interface (HCI) for virtual sport activities. In the proposed technique, a graphical 3D human model is first constructed. Its projection on a virtual image plane is then used to match the silhouettes obtained from the image sequence. By iteratively adjusting the 3D pose of the graphical 3D model with the physical and anatomic constraints of human motion, the human pose and the associate 3D motion parameters can be uniquely identified. Experimental results are presented with the real scene images.

In this paper, we propose a self-localization and path-planning method for mobile robot navigation. An omnidirectional camera and infrared sensors are used to extract the landmarks information of the environment. Due to the large field of view of the omnidirectional camera, the mobile robot can capture the rich information of the environment. The landmark features are detected and extracted from the omnidirectional video camera, so the robot is able to navigate in the environment automatically to learn the localization information and avoid obstacles by using infrared sensors. The robot system can then use the localization information to plan a shortest path to visit some particular locations prespecified by the user.

This paper presents a practical optimization procedure for object detection and recognition algorithms. It is suitable for object recognition using a catadioptric omnidirectional vision system mounted on a mobile robot. We use the SIFT descriptor to obtain image features of the objects and the environment. First, sample object images are given for training and optimization procedures. Bayesian classification is used to train various test objects based on different SIFT vectors. The system selects the features based on the k-means group to predict the possible object from the candidate regions of the images. It is thus able to detect the object with arbitrary shape without the 3D information. The feature optimization procedure makes the object features more stable for recognition and classification. Experimental results are presented for real scene images captured by a catadioptric omni-vision camera.

This paper presents a depth recovery technique based on the maximum defocus blur associated with a camera focus setting. The depth-blur relation is formulated by a mathematical model and verified by defocus calibration. Image intensity histogram analysis is used to identify the blur extent. Different from the existing depth from defocus approaches, our method is capable of depth recovery using a single image, possibly up to scale. Experiments on real scene images have demonstrated the feasibility of the proposed method for depth recovery.

In this paper we propose a 3D reconstruction algorithm by combining shape from silhouette with stereo. Visual hull of the object is first derived from multi-view silhouette images. Pairwise stereo matching for shape refinement is then accomplished using the best viewable images. Based on the reduced correspondence searching range constrained by contact points and bounding edges, significant improvement of visual hull is possible even if the number of cameras is limited. Experimental results are presented for both synthetic data and real scene images.

Motion blur is the result when the camera shutter remains open for an extended period of time and a relative motion between camera and object occurs. Most research on this type of image degradation is focused on motion blur removal. In this work, we propose a novel approach for vehicle speed detection based on motion blurred images. The motion blur parameters are first estimated from the acquired images and then used to detect the speed of the moving object in the scene. We have established a link between the motion blur information of a 2D image and the speed information of a moving object. Experimental results are presented for both indoor environments and outdoor vehicle speed detection.

A desktop vision system is presented for complete 3D model acquisition. It is fast, low-cost, and accurate. Partial 3D shapes and texture information are acquired from multiple viewing directions using rotational stereo and shape from focus (SFF). The resulting range images are registered to a common coordinate system and a surface representation is created for each range image. The resulting surfaces are integrated using a new algorithm named Region-of-Construction. Unlike previous approaches, the Region-of-Construction algorithm directly exploits the structure of the raw range images. The algorithm determines regions in range images corresponding to non-redundant surfaces which can be stitched along the boundaries to construct the complete 3D surface model. The algorithm is computationally efficient and less sensitive to registration error. It also has the ability to construct complete 3D models of complex objects with holes. A photo realistic 3D model is obtained by mapping texture information onto the complete surface model representing 3D shape. Experimental results for several real objects are presented.