Denoising in optical coherence tomography volumes for improved 3D visualization.

Optical coherence tomography (OCT) has already become one of the most important diagnostic tools in different fields of medicine, as well as in various industrial applications. The most important characteristic of OCT is its high resolution, both in depth and the transverse direction. Together with the information on the tissue density, OCT offers highly precise information on tissue geometry. However, the detectability of small and low-intensity features in OCT scans is limited by the presence of speckle noise. In this paper we present a new volumetric method for noise removal in OCT volumes, which aims at improving the quality of rendered 3D volumes. In order to remove noise uniformly, while preserving important details, the proposed algorithm simultaneously observes the estimated amounts of noise and the sharpness measure, and iteratively enhances the volume until it reaches the required quality. We evaluate the proposed method using four quality measures as well as visually, by evaluating the visualization of OCT volumes on an auto-stereoscopic 3D screen. The results show that the proposed method outperforms reference methods both visually and in terms of objective measures.

Adaptive point cloud acquisition and upsampling for automotive lidar.

One of the crucial factors in achieving a higher level of autonomy of self-driving vehicles is a sensor capable of acquiring accurate and robust information about the environment and other participants in traffic. In the past few decades, various types of sensors have been used for this purpose, such as cameras registering visible, near-infrared, and thermal parts of the spectrum, as well as radars, ultrasonic sensors, and lidar. Due to their high range, accuracy, and robustness, lidars are gaining popularity in numerous applications. However, in many cases, their spatial resolution does not meet the requirements of the application. To solve this problem, we propose a strategy for better utilization of the available points. In particular, we propose an adaptive paradigm that scans the objects of interest with increased resolution, while the background is scanned using a lower point density. Initial region proposals are generated using an object detector that relies on an auxiliary camera. Such a strategy improves the quality of the representation of the object, while retaining the total number of projected points. The proposed method shows improvements compared to regular sampling in terms of the quality of upsampled point clouds.

High-Dynamic-Range Tone Mapping in Intelligent Automotive Systems.

Intelligent driver assistance systems are becoming increasingly popular in modern passenger vehicles. A crucial component of intelligent vehicles is the ability to detect vulnerable road users (VRUs) for an early and safe response. However, standard imaging sensors perform poorly in conditions of strong illumination contrast, such as approaching a tunnel or at night, due to their dynamic range limitations. In this paper, we focus on the use of high-dynamic-range (HDR) imaging sensors in vehicle perception systems and the subsequent need for tone mapping of the acquired data into a standard 8-bit representation. To our knowledge, no previous studies have evaluated the impact of tone mapping on object detection performance. We investigate the potential for optimizing HDR tone mapping to achieve a natural image appearance while facilitating object detection of state-of-the-art detectors designed for standard dynamic range (SDR) images. Our proposed approach relies on a lightweight convolutional neural network (CNN) that tone maps HDR video frames into a standard 8-bit representation. We introduce a novel training approach called detection-informed tone mapping (DI-TM) and evaluate its performance with respect to its effectiveness and robustness in various scene conditions, as well as its performance relative to an existing state-of-the-art tone mapping method. The results show that the proposed DI-TM method achieves the best results in terms of detection performance metrics in challenging dynamic range conditions, while both methods perform well in typical, non-challenging conditions. In challenging conditions, our method improves the detection F2 score by 13%. Compared to SDR images, the increase in F2 score is 49%.

Deep Learning Tone-Mapping and Demosaicing for Automotive Vision Systems.

High dynamic range (HDR) imaging technology is increasingly being used in automated driving systems (ADS) for improving the safety of traffic participants in scenes with strong differences in illumination. Therefore, a combination of HDR video, that is video with details in all illumination regimes, and (HDR) object perception techniques that can deal with this variety in illumination is highly desirable. Although progress has been made in both HDR imaging solutions and object detection algorithms in the recent years, they have progressed independently of each other. This has led to a situation in which object detection algorithms are typically designed and constantly improved to operate on 8 bit per channel content. This makes these algorithms not ideally suited for use in HDR data processing, which natively encodes to a higher bit-depth (12 bits/16 bits per channel). In this paper, we present and evaluate two novel convolutional neural network (CNN) architectures that intelligently convert high bit depth HDR images into 8-bit images. We attempt to optimize reconstruction quality by focusing on ADS object detection quality. The first research novelty is to jointly perform tone-mapping with demosaicing by additionally successfully suppressing noise and demosaicing artifacts. The first CNN performs tone-mapping with noise suppression on a full-color HDR input, while the second performs joint demosaicing and tone-mapping with noise suppression on a raw HDR input. The focus is to increase the detectability of traffic-related objects in the reconstructed 8-bit content, while ensuring that the realism of the standard dynamic range (SDR) content in diverse conditions is preserved. The second research novelty is that for the first time, to the best of our knowledge, a thorough comparative analysis against the state-of-the-art tone-mapping and demosaicing methods is performed with respect to ADS object detection accuracy on traffic-related content that abounds with diverse challenging (i.e., boundary cases) scenes. The evaluation results show that the two proposed networks have better performance in object detection accuracy and image quality, than both SDR content and content obtained with the state-of-the-art tone-mapping and demosaicing algorithms.

Improving Turn Movement Count Using Cooperative Feedback.

In this paper, we propose a new cooperative method that improves the accuracy of Turn Movement Count (TMC) under challenging conditions by introducing contextual observations from the surrounding areas. The proposed method focuses on the correct identification of the movements in conditions where current methods have difficulties. Existing vision-based TMC systems are limited under heavy traffic conditions. The main problems for most existing methods are occlusions between vehicles that prevent the correct detection and tracking of the vehicles through the entire intersection and the assessment of the vehicle's entry and exit points, incorrectly assigning the movement. The proposed method intends to overcome this incapability by sharing information with other observation systems located at neighboring intersections. Shared information is used in a cooperative scheme to infer the missing data, thereby improving the assessment that would otherwise not be counted or miscounted. Experimental evaluation of the system shows a clear improvement over related reference methods.

General Image Fusion for an Arbitrary Number of Inputs Using Convolutional Neural Networks.

In this paper, we propose a unified and flexible framework for general image fusion tasks, including multi-exposure image fusion, multi-focus image fusion, infrared/visible image fusion, and multi-modality medical image fusion. Unlike other deep learning-based image fusion methods applied to a fixed number of input sources (normally two inputs), the proposed framework can simultaneously handle an arbitrary number of inputs. Specifically, we use the symmetrical function (e.g., Max-pooling) to extract the most significant features from all the input images, which are then fused with the respective features from each input source. This symmetry function enables permutation-invariance of the network, which means the network can successfully extract and fuse the saliency features of each image without needing to remember the input order of the inputs. The property of permutation-invariance also brings convenience for the network during inference with unfixed inputs. To handle multiple image fusion tasks with one unified framework, we adopt continual learning based on Elastic Weight Consolidation (EWC) for different fusion tasks. Subjective and objective experiments on several public datasets demonstrate that the proposed method outperforms state-of-the-art methods on multiple image fusion tasks.

An Unsupervised Transfer Learning Framework for Visible-Thermal Pedestrian Detection.

Dual cameras with visible-thermal multispectral pairs provide both visual and thermal appearance, thereby enabling detecting pedestrians around the clock in various conditions and applications, including autonomous driving and intelligent transportation systems. However, due to the greatly varying real-world scenarios, the performance of a detector trained on a source dataset might change dramatically when evaluated on another dataset. A large amount of training data is often necessary to guarantee the detection performance in a new scenario. Typically, human annotators need to conduct the data labeling work, which is time-consuming, labor-intensive and unscalable. To overcome the problem, we propose a novel unsupervised transfer learning framework for multispectral pedestrian detection, which adapts a multispectral pedestrian detector to the target domain based on pseudo training labels. In particular, auxiliary detectors are utilized and different label fusion strategies are introduced according to the estimated environmental illumination level. Intermediate domain images are generated by translating the source images to mimic the target ones, acting as a better starting point for the parameter update of the pedestrian detector. The experimental results on the KAIST and FLIR ADAS datasets demonstrate that the proposed method achieves new state-of-the-art performance without any manual training annotations on the target data.

Region-Based CNN for Anomaly Detection in PV Power Plants Using Aerial Imagery.

Today, solar energy is taking an increasing share of the total energy mix. Unfortunately, many operational photovoltaic plants suffer from a plenitude of defects resulting in non-negligible power loss. The latter highly impacts the overall performance of the PV site; therefore, operators need to regularly inspect their solar parks for anomalies in order to prevent severe performance drops. As this operation is naturally labor-intensive and costly, we present in this paper a novel system for improved PV diagnostics using drone-based imagery. Our solution consists of three main steps. The first step locates the solar panels within the image. The second step detects the anomalies within the solar panels. The final step identifies the root cause of the anomaly. In this paper, we mainly focus on the second step comprising the detection of anomalies within solar panels, which is done using a region-based convolutional neural network (CNN). Experiments on six different PV sites with different specifications and a variety of defects demonstrate that our anomaly detector achieves a true positive rate or recall of more than 90% for a false positive rate of around 2% to 3% tested on a dataset containing nearly 9000 solar panels. Compared to the best state-of-the-art methods, the experiments revealed that we achieve a slightly higher true positive rate for a substantially lower false positive rate, while tested on a more realistic dataset.

Deep HDR Deghosting by Motion-Attention Fusion Network.

Multi-exposure image fusion (MEF) methods for high dynamic range (HDR) imaging suffer from ghosting artifacts when dealing with moving objects in dynamic scenes. The state-of-the-art methods use optical flow to align low dynamic range (LDR) images before merging, introducing distortion into the aligned LDR images from inaccurate motion estimation due to large motion and occlusion. In place of pre-alignment, attention-based methods calculate the correlation between the reference LDR image and non-reference LDR images, thus excluding misaligned regions in LDR images. Nevertheless, they also exclude the saturated details at the same time. Taking advantage of both the alignment and attention-based methods, we propose an efficient Deep HDR Deghosting Fusion Network (DDFNet) guided by optical flow and image correlation attentions. Specifically, the DDFNet estimates the optical flow of the LDR images by a motion estimation module and encodes that optical flow as a flow feature. Additionally, it extracts correlation features between the reference LDR and other non-reference LDR images. The optical flow and correlation features are employed to adaptably combine information from LDR inputs in an attention-based fusion module. Following the merging of features, a decoder composed of Dense Networks reconstructs the HDR image without ghosting. Experimental results indicate that the proposed DDFNet achieves state-of-the-art image fusion performance on different public datasets.

Probabilistic Fusion for Pedestrian Detection from Thermal and Colour Images.

Pedestrian detection is an important research domain due to its relevance for autonomous and assisted driving, as well as its applications in security and industrial automation. Often, more than one type of sensor is used to cover a broader range of operating conditions than a single-sensor system would allow. However, it remains difficult to make pedestrian detection systems perform well in highly dynamic environments, often requiring extensive retraining of the algorithms for specific conditions to reach satisfactory accuracy, which, in turn, requires large, annotated datasets captured in these conditions. In this paper, we propose a probabilistic decision-level sensor fusion method based on naive Bayes to improve the efficiency of the system by combining the output of available pedestrian detectors for colour and thermal images without retraining. The results in this paper, obtained through long-term experiments, demonstrate the efficacy of our technique, its ability to work with non-registered images, and its adaptability to cope with situations when one of the sensors fails. The results also show that our proposed technique improves the overall accuracy of the system and could be very useful in several applications.

Encoding Stability into Laser Powder Bed Fusion Monitoring Using Temporal Features and Pore Density Modelling.

In laser powder bed fusion (LPBF), melt pool instability can lead to the development of pores in printed parts, reducing the part's structural strength. While camera-based monitoring systems have been introduced to improve melt pool stability, these systems only measure melt pool stability in limited, indirect ways. We propose that melt pool stability can be improved by explicitly encoding stability into LPBF monitoring systems through the use of temporal features and pore density modelling. We introduce the temporal features, in the form of temporal variances of common LPBF monitoring features (e.g., melt pool area, intensity), to explicitly quantify printing stability. Furthermore, we introduce a neural network model trained to link these video features directly to pore densities estimated from the CT scans of previously printed parts. This model aims to reduce the number of online printer interventions to only those that are required to avoid porosity. These contributions are then implemented in a full LPBF monitoring system and tested on prints using 316L stainless steel. Results showed that our explicit stability quantification improved the correlation between our predicted pore densities and true pore densities by up to 42%.

Depth Completion and Super-Resolution with Arbitrary Scale Factors for Indoor Scenes.

Depth sensing has improved rapidly in recent years, which allows for structural information to be utilized in various applications, such as virtual reality, scene and object recognition, view synthesis, and 3D reconstruction. Due to the limitations of the current generation of depth sensors, the resolution of depth maps is often still much lower than the resolution of color images. This hinders applications, such as view synthesis or 3D reconstruction, from providing high-quality results. Therefore, super-resolution, which allows for the upscaling of depth maps while still retaining sharpness, has recently drawn much attention in the deep learning community. However, state-of-the-art deep learning methods are typically designed and trained to handle a fixed set of integer-scale factors. Moreover, the raw depth map collected by the depth sensor usually has many depth data missing or misestimated values along the edges and corners of observed objects. In this work, we propose a novel deep learning network for both depth completion and depth super-resolution with arbitrary scale factors. The experimental results on the Middlebury stereo, NYUv2, and Matterport3D datasets demonstrate that the proposed method can outperform state-of-the-art methods.

Discovering Human Activities from Binary Data in Smart Homes.

With the rapid development in sensing technology, data mining, and machine learning fields for human health monitoring, it became possible to enable monitoring of personal motion and vital signs in a manner that minimizes the disruption of an individual's daily routine and assist individuals with difficulties to live independently at home. A primary difficulty that researchers confront is acquiring an adequate amount of labeled data for model training and validation purposes. Therefore, activity discovery handles the problem that activity labels are not available using approaches based on sequence mining and clustering. In this paper, we introduce an unsupervised method for discovering activities from a network of motion detectors in a smart home setting. First, we present an intra-day clustering algorithm to find frequent sequential patterns within a day. As a second step, we present an inter-day clustering algorithm to find the common frequent patterns between days. Furthermore, we refine the patterns to have more compressed and defined cluster characterizations. Finally, we track the occurrences of various regular routines to monitor the functional health in an individual's patterns and lifestyle. We evaluate our methods on two public data sets captured in real-life settings from two apartments during seven-month and three-month periods.

Efficient Training Procedures for Multi-Spectra Demosaicing.

The simultaneous acquisition of multi-spectral images on a single sensor can be efficiently performed by single shot capture using a mutli-spectral filter array. This paper focused on the demosaicing of color and near-infrared bands and relied on a convolutional neural network (CNN). To train the deep learning model robustly and accurately, it is necessary to provide enough training data, with sufficient variability. We focused on the design of an efficient training procedure by discovering an optimal training dataset. We propose two data selection strategies, motivated by slightly different concepts. The general term that will be used for the proposed models trained using data selection is data selection-based multi-spectral demosaicing (DSMD). The first idea is clustering-based data selection (DSMD-C), with the goal to discover a representative subset with a high variance so as to train a robust model. The second is an adaptive-based data selection (DSMD-A), a self-guided approach that selects new data based on the current model accuracy. We performed a controlled experimental evaluation of the proposed training strategies and the results show that a careful selection of data does benefit the speed and accuracy of training. We are still able to achieve high reconstruction accuracy with a lightweight model.

Cooperative Multi-Sensor Tracking of Vulnerable Road Users in the Presence of Missing Detections.

This paper presents a vulnerable road user (VRU) tracking algorithm capable of handling noisy and missing detections from heterogeneous sensors. We propose a cooperative fusion algorithm for matching and reinforcing of radar and camera detections using their proximity and positional uncertainty. The belief in the existence and position of objects is then maximized by temporal integration of fused detections by a multi-object tracker. By switching between observation models, the tracker adapts to the detection noise characteristics making it robust to individual sensor failures. The main novelty of this paper is an improved imputation sampling function for updating the state when detections are missing. The proposed function uses a likelihood without association that is conditioned on the sensor information instead of the sensor model. The benefits of the proposed solution are two-fold: firstly, particle updates become computationally tractable and secondly, the problem of imputing samples from a state which is predicted without an associated detection is bypassed. Experimental evaluation shows a significant improvement in both detection and tracking performance over multiple control algorithms. In low light situations, the cooperative fusion outperforms intermediate fusion by as much as 30%, while increases in tracking performance are most significant in complex traffic scenes.

Camera-Based System for Drafting Detection While Cycling.

Drafting involves cycling so close behind another person that wind resistance is significantly reduced, which is illegal during most long distance and several short distance triathlon and duathlon events. In this paper, a proof of concept for a drafting detection system based on computer vision is proposed. After detecting and tracking a bicycle through the various scenes, the distance to this object is estimated through computational geometry. The probability of drafting is then determined through statistical analysis of subsequent measurements over an extended period of time. These algorithms are tested using a static recording and a recording that simulates a race situation with ground truth distances obtained from a Light Detection And Ranging (LiDAR) system. The most accurate developed distance estimation method yields an average error of 0 . 46 m in our test scenario. When sampling the distances at periods of 1 or 2 s, simulations demonstrate that a drafting violation is detected quickly for cyclists riding at 2 m or more below the limit, while generally avoiding false positives during the race-like test set-up and five hour race simulations.

Deep Visible and Thermal Image Fusion for Enhanced Pedestrian Visibility.

Reliable vision in challenging illumination conditions is one of the crucial requirements of future autonomous automotive systems. In the last decade, thermal cameras have become more easily accessible to a larger number of researchers. This has resulted in numerous studies which confirmed the benefits of the thermal cameras in limited visibility conditions. In this paper, we propose a learning-based method for visible and thermal image fusion that focuses on generating fused images with high visual similarity to regular truecolor (red-green-blue or RGB) images, while introducing new informative details in pedestrian regions. The goal is to create natural, intuitive images that would be more informative than a regular RGB camera to a human driver in challenging visibility conditions. The main novelty of this paper is the idea to rely on two types of objective functions for optimization: a similarity metric between the RGB input and the fused output to achieve natural image appearance; and an auxiliary pedestrian detection error to help defining relevant features of the human appearance and blending them into the output. We train a convolutional neural network using image samples from variable conditions (day and night) so that the network learns the appearance of humans in the different modalities and creates more robust results applicable in realistic situations. Our experiments show that the visibility of pedestrians is noticeably improved especially in dark regions and at night. Compared to existing methods we can better learn context and define fusion rules that focus on the pedestrian appearance, while that is not guaranteed with methods that focus on low-level image quality metrics.

Behavioral Pedestrian Tracking Using a Camera and LiDAR Sensors on a Moving Vehicle.

In this paper, we present a novel 2D⁻3D pedestrian tracker designed for applications in autonomous vehicles. The system operates on a tracking by detection principle and can track multiple pedestrians in complex urban traffic situations. By using a behavioral motion model and a non-parametric distribution as state model, we are able to accurately track unpredictable pedestrian motion in the presence of heavy occlusion. Tracking is performed independently, on the image and ground plane, in global, motion compensated coordinates. We employ Camera and LiDAR data fusion to solve the association problem where the optimal solution is found by matching 2D and 3D detections to tracks using a joint log-likelihood observation model. Each 2D⁻3D particle filter then updates their state from associated observations and a behavioral motion model. Each particle moves independently following the pedestrian motion parameters which we learned offline from an annotated training dataset. Temporal stability of the state variables is achieved by modeling each track as a Markov Decision Process with probabilistic state transition properties. A novel track management system then handles high level actions such as track creation, deletion and interaction. Using a probabilistic track score the track manager can cull false and ambiguous detections while updating tracks with detections from actual pedestrians. Our system is implemented on a GPU and exploits the massively parallelizable nature of particle filters. Due to the Markovian nature of our track representation, the system achieves real-time performance operating with a minimal memory footprint. Exhaustive and independent evaluation of our tracker was performed by the KITTI benchmark server, where it was tested against a wide variety of unknown pedestrian tracking situations. On this realistic benchmark, we outperform all published pedestrian trackers in a multitude of tracking metrics.

Automatic Multi-Camera Extrinsic Parameter Calibration Based on Pedestrian Torsors.

Extrinsic camera calibration is essential for any computer vision task in a camera network. Typically, researchers place a calibration object in the scene to calibrate all the cameras in a camera network. However, when installing cameras in the field, this approach can be costly and impractical, especially when recalibration is needed. This paper proposes a novel, accurate and fully automatic extrinsic calibration framework for camera networks with partially overlapping views. The proposed method considers the pedestrians in the observed scene as the calibration objects and analyzes the pedestrian tracks to obtain extrinsic parameters. Compared to the state of the art, the new method is fully automatic and robust in various environments. Our method detect human poses in the camera images and then models walking persons as vertical sticks. We apply a brute-force method to determines the correspondence between persons in multiple camera images. This information along with 3D estimated locations of the top and the bottom of the pedestrians are then used to compute the extrinsic calibration matrices. We also propose a novel method to calibrate the camera network by only using the top and centerline of the person when the bottom of the person is not available in heavily occluded scenes. We verified the robustness of the method in different camera setups and for both single and multiple walking people. The results show that the triangulation error of a few centimeters can be obtained. Typically, it requires less than one minute of observing the walking people to reach this accuracy in controlled environments. It also just takes a few minutes to collect enough data for the calibration in uncontrolled environments. Our proposed method can perform well in various situations such as multi-person, occlusions, or even at real intersections on the street.

Robust Pan/Tilt Compensation for Foreground-Background Segmentation.

In this paper, we describe a robust method for compensating the panning and tilting motion of a camera, applied to foreground-background segmentation. First, the necessary internal camera parameters are determined through feature-point extraction and tracking. From these parameters, two motion models for points in the image plane are established. The first model assumes a fixed tilt angle, whereas the second model allows simultaneous pan and tilt. At runtime, these models are used to compensate for the motion of the camera in the background model. We will show that these methods provide a robust compensation mechanism and improve the foreground masks of an otherwise state-of-the-art unsupervised foreground-background segmentation method. The resulting algorithm is always able to obtain F 1 scores above 80 % on every daytime video in our test set when a minimal number of only eight feature matches are used to determine the background compensation, whereas the standard approaches need significantly more feature matches to produce similar results.