Loop closing is a fundamental part of simultaneous localization and mapping (SLAM) for autonomous mobile systems. In the field of visual SLAM, bag of words (BoW) has achieved great success in loop closure. The BoW features for loop searching can also be used in the subsequent 6-DoF loop correction. However, for 3D LiDAR SLAM, the state-of-the-art methods may fail to effectively recognize the loop in real time, and usually cannot correct the full 6-DoF loop pose. To address this limitation, we present a novel B ag o f W ords for real-time loop closing in 3D LiDAR SLAM, called BoW3D. Our method not only efficiently recognizes the revisited loop places, but also corrects the full 6-DoF loop pose in real time. BoW3D builds the bag of words based on the 3D LiDAR feature LinK3D, which is efficient, pose-invariant and can be used for accurate point-to-point matching. We furthermore embed our proposed method into 3D LiDAR odometry system to evaluate loop closing performance. We test our method on public dataset, and compare it against other state-of-the-art algorithms. BoW3D shows better performance in terms of $F_{1}$ max and extended precision scores on most scenes. It is noticeable that BoW3D takes an average of 48 ms to recognize and correct the loops on KITTI 00 (includes 4K+ 64-ray LiDAR scans), when executed on a notebook with an Intel Core i7 @2.2 GHz processor.

Feature extraction and matching are the basic parts of many robotic vision tasks, such as 2D or 3D object detection, recognition, and registration. As is known, 2D feature extraction and matching have already achieved great success. Unfortunately, in the field of 3D, the current methods may fail to support the extensive application of 3D LiDAR sensors in robotic vision tasks due to their poor descriptiveness and inefficiency. To address this limitation, we propose a novel 3D feature representation method: Lin ear K eypoints representation for 3D LiDAR point cloud, called LinK3D. The novelty of LinK3D lies in that it fully considers the characteristics (such as the sparsity and complexity) of LiDAR point clouds and represents the keypoint with its robust neighbor keypoints, which provide strong constraints in the description of the keypoint. The proposed LinK3D has been evaluated on three public datasets, and the experimental results show that our method achieves great matching performance. More importantly, LinK3D also shows excellent real-time performance, faster than the sensor frame rate at 10 Hz of a typical rotating LiDAR sensor. LinK3D only takes an average of $\bm {30}$ milliseconds to extract features from the point cloud collected by a 64-beam LiDAR and takes merely about $\bm {20}$ milliseconds to match two LiDAR scans when executed on a computer with an Intel Core i7 processor. Moreover, our method can be extended to LiDAR odometry task, and shows good scalability.

This paper presents a novel approach to ECG signal filtering and classification. Unlike the traditional techniques which aim at collecting and processing the ECG signals with the patient being still, lying in bed in hospitals, our proposed algorithm is intentionally designed for monitoring and classifying the patient’s ECG signals in the free-living environment. The patients are equipped with wearable ambulatory devices the whole day, which facilitates the real-time heart attack detection. In ECG preprocessing, an integral-coefficient-band-stop (ICBS) filter is applied, which omits time-consuming floating-point computations. In addition, two-layered Hidden Markov Models (HMMs) are applied to achieve ECG feature extraction and classification. The periodic ECG waveforms are segmented into ISO intervals, P subwave, QRS complex and T subwave respectively in the first HMM layer where expert-annotation assisted Baum-Welch algorithm is utilized in HMM modeling. Then the corresponding interval features are selected and applied to categorize the ECG into normal type or abnormal type (PVC, APC) in the second HMM layer. For verifying the effectiveness of our algorithm on abnormal signal detection, we have developed an ECG body sensor network (BSN) platform, whereby real-time ECG signals are collected, transmitted, displayed and the corresponding classification outcomes are deduced and shown on the BSN screen.

Abnormal behavior has become the leading cause of injuries among the elderly in the modern society. Elderly anomaly is a widespread concern in both academic and industrial fields. Detection of abnormal behaviors using deep learning techniques has been pervasively investigated due to the flourishing of powerful edge computing and the decreasing costs of consumer electronics, such as commercial vision sensors. This paper intentionally designs an innovative architecture to detect the elderly typical abnormal behaviors: falls and tumbles, aggression and wandering, using the consumer network cameras, which are configured in the residential areas, healthcare centers, to name a few. Different from the conventional abnormal behavior detection methods, which suffer the issues of improper model generalization and lack of the spatial and temporal coherence, we design a deep learning model that robustly extracts the skeleton joints, detects and classifies the abnormal behaviors while considering the spatial and temporal context. It should be noted that the images are captured with the network cameras rigidly fixed at the typical elderly activity spots. The group of captured images will be fed to the local server, equipped with GPUs, for online monitoring and alarming. Our method is extensively evaluated on the elderly anomaly detection platform. It achieves the competitive performance with mAP greater than 85%.

This article designs a visual surveillance framework for human fall detection. In order to solve the conventional issues in fall detection, such as unsatisfactory feature generalization, low recall rates, and large computational time, we design a model that incorporates the deep convolutional neural network and the aggregated heuristic visual features in detecting the occurrence of falls. First, the convolutional neural network (Openpose model) is utilized to extract human skeleton in the image. Second, the hand-crafted spatial features, such as the angle of human shank inclination, are aggregated to determine the fall presence. It should be noticed that our fall detection method has been integrated to healthcare Internet of Things (IoT) video surveillance architecture, which has multiple graphic processing unit groups to perform real-time monitoring and alarming for the elderly in need. The experimental results prove that our method is able to accurately distinguish fall and nonfall activities with a competitive false-alarm rate.

Autonomous underwater vehicles (AUVs) equipped with online visual inspection systems can detect underwater targets during underwater operations, which is of great significance to subsea exploration. However, the undersea scene has some instinctive challenging problems, such as poor lighting conditions, sediment burial, and marine biofouling mimicry, which makes it difficult for traditional target detection algorithms to achieve online, reliable, and accurate detection of underwater targets. To solve the above issues, this paper proposes a real-time object detection algorithm for underwater targets based on a lightweight convolutional neural network model. To improve the imaging quality of underwater images, contrast limited adaptive histogram equalization with the fused multicolor space (FCLAHE) model is designed to enhance the image quality of underwater targets. Afterwards, a spindle-shaped backbone network is designed. The inverted residual block and group convolutions are used to extract depth features to ensure the target detection accuracy on one hand and to reduce the model parameter volume on the other hand under complex scenarios. Through extensive experiments, the precision, recall, and mAP of the proposed algorithm reached 91.2%, 90.1%, and 88.3%, respectively. It is also noticeable that the proposed method has been integrated into the embedded GPU platform and deployed in the AUV system in the practical scenarios. The average computational time is 0.053s, which satisfies the requirements of real-time object detection.