Behavior of bats at wind turbines

Wind turbines are causing unprecedented numbers of bat fatalities. Many fatalities involve tree-roosting bats, but reasons for this higher susceptibility remain unknown. To better understand behaviors associated with risk, we monitored bats at three experimentally manipulated wind turbines in Indiana, United States, from July 29 to October 1, 2012, using thermal cameras and other methods. We observed bats on 993 occasions and saw many behaviors, including close approaches, flight loops and dives, hovering, and chases. Most bats altered course toward turbines during observation. Based on these new observations, we tested the hypotheses that wind speed and blade rotation speed influenced the way that bats interacted with turbines. We found that bats were detected more frequently at lower wind speeds and typically approached turbines on the leeward (downwind) side. The proportion of leeward approaches increased with wind speed when blades were prevented from turning, yet decreased when blades could turn. Bats were observed more frequently at turbines on moonlit nights. Taken together, these observations suggest that bats may orient toward turbines by sensing air currents and using vision, and that air turbulence caused by fast-moving blades creates conditions that are less attractive to bats passing in close proximity. Tree bats may respond to streams of air flowing downwind from trees at night while searching for roosts, conspecifics, and nocturnal insect prey that could accumulate in such flows. Fatalities of tree bats at turbines may be the consequence of behaviors that evolved to provide selective advantages when elicited by tall trees, but are now maladaptive when elicited by wind turbines.

Bridging the gap: from biometrics to forensics

Biometric recognition, or simply biometrics, refers to automated recognition of individuals based on their behavioural and biological characteristics. The success of fingerprints in forensic science and law enforcement applications, coupled with growing concerns related to border control, financial fraud and cyber security, has generated a huge interest in using fingerprints, as well as other biological traits, for automated person recognition. It is, therefore, not surprising to see biometrics permeating various segments of our society. Applications include smartphone security, mobile payment, border crossing, national civil registry and access to restricted facilities. Despite these successful deployments in various fields, there are several existing challenges and new opportunities for person recognition using biometrics. In particular, when biometric data is acquired in an unconstrained environment or if the subject is uncooperative, the quality of the ensuing biometric data may not be amenable for automated person recognition. This is particularly true in crime-scene investigations, where the biological evidence gleaned from a scene may be of poor quality. In this article, we first discuss how biometrics evolved from forensic science and how its focus is shifting back to its origin in order to address some challenging problems. Next, we enumerate the similarities and differences between biometrics and forensics. We then present some applications where the principles of biometrics are being successfully leveraged into forensics in order to solve critical problems in the law enforcement domain. Finally, we discuss new collaborative opportunities for researchers in biometrics and forensics, in order to address hitherto unsolved problems that can benefit society at large.

Cyber surveillance for flood disasters

Regional heavy rainfall is usually caused by the influence of extreme weather conditions. Instant heavy rainfall often results in the flooding of rivers and the neighboring low-lying areas, which is responsible for a large number of casualties and considerable property loss. The existing precipitation forecast systems mostly focus on the analysis and forecast of large-scale areas but do not provide precise instant automatic monitoring and alert feedback for individual river areas and sections. Therefore, in this paper, we propose an easy method to automatically monitor the flood object of a specific area, based on the currently widely used remote cyber surveillance systems and image processing methods, in order to obtain instant flooding and waterlogging event feedback. The intrusion detection mode of these surveillance systems is used in this study, wherein a flood is considered a possible invasion object. Through the detection and verification of flood objects, automatic flood risk-level monitoring of specific individual river segments, as well as the automatic urban inundation detection, has become possible. The proposed method can better meet the practical needs of disaster prevention than the method of large-area forecasting. It also has several other advantages, such as flexibility in location selection, no requirement of a standard water-level ruler, and a relatively large field of view, when compared with the traditional water-level measurements using video screens. The results can offer prompt reference for appropriate disaster warning actions in small areas, making them more accurate and effective.

Towards the sustainable development of smart cities through mass video surveillance: A response to the COVID-19 pandemic

Sustainable smart city initiatives around the world have recently had great impact on the lives of citizens and brought significant changes to society. More precisely, data-driven smart applications that efficiently manage sparse resources are offering a futuristic vision of smart, efficient, and secure city operations. However, the ongoing COVID-19 pandemic has revealed the limitations of existing smart city deployment; hence; the development of systems and architectures capable of providing fast and effective mechanisms to limit further spread of the virus has become paramount. An active surveillance system capable of monitoring and enforcing social distancing between people can effectively slow the spread of this deadly virus. In this paper, we propose a data-driven deep learning-based framework for the sustainable development of a smart city, offering a timely response to combat the COVID-19 pandemic through mass video surveillance. To implementing social distancing monitoring, we used three deep learning-based real-time object detection models for the detection of people in videos captured with a monocular camera. We validated the performance of our system using a real-world video surveillance dataset for effective deployment.

An early fire detection algorithm using IP cameras

The presence of smoke is the first symptom of fire; therefore to achieve early fire detection, accurate and quick estimation of the presence of smoke is very important. In this paper we propose an algorithm to detect the presence of smoke using video sequences captured by Internet Protocol (IP) cameras, in which important features of smoke, such as color, motion and growth properties are employed. For an efficient smoke detection in the IP camera platform, a detection algorithm must operate directly in the Discrete Cosine Transform (DCT) domain to reduce computational cost, avoiding a complete decoding process required for algorithms that operate in spatial domain. In the proposed algorithm the DCT Inter-transformation technique is used to increase the detection accuracy without inverse DCT operation. In the proposed scheme, firstly the candidate smoke regions are estimated using motion and color smoke properties; next using morphological operations the noise is reduced. Finally the growth properties of the candidate smoke regions are furthermore analyzed through time using the connected component labeling technique. Evaluation results show that a feasible smoke detection method with false negative and false positive error rates approximately equal to 4% and 2%, respectively, is obtained.

Parallax-Robust Surveillance Video Stitching

This paper presents a parallax-robust video stitching technique for timely synchronized surveillance video. An efficient two-stage video stitching procedure is proposed in this paper to build wide Field-of-View (FOV) videos for surveillance applications. In the stitching model calculation stage, we develop a layered warping algorithm to align the background scenes, which is location-dependent and turned out to be more robust to parallax than the traditional global projective warping methods. On the selective seam updating stage, we propose a change-detection based optimal seam selection approach to avert ghosting and artifacts caused by moving foregrounds. Experimental results demonstrate that our procedure can efficiently stitch multi-view videos into a wide FOV video output without ghosting and noticeable seams.

FPGA implementation for real-time background subtraction based on Horprasert model

Background subtraction is considered the first processing stage in video surveillance systems, and consists of determining objects in movement in a scene captured by a static camera. It is an intensive task with a high computational cost. This work proposes an embedded novel architecture on FPGA which is able to extract the background on resource-limited environments and offers low degradation (produced because of the hardware-friendly model modification). In addition, the original model is extended in order to detect shadows and improve the quality of the segmentation of the moving objects. We have analyzed the resource consumption and performance in Spartan3 Xilinx FPGAs and compared to others works available on the literature, showing that the current architecture is a good trade-off in terms of accuracy, performance and resources utilization. With less than a 65% of the resources utilization of a XC3SD3400 Spartan-3A low-cost family FPGA, the system achieves a frequency of 66.5 MHz reaching 32.8 fps with resolution 1,024 × 1,024 pixels, and an estimated power consumption of 5.76 W.

Improving Video Segmentation by Fusing Depth Cues and the Visual Background Extractor (ViBe) Algorithm

Depth-sensing technology has led to broad applications of inexpensive depth cameras that can capture human motion and scenes in three-dimensional space. Background subtraction algorithms can be improved by fusing color and depth cues, thereby allowing many issues encountered in classical color segmentation to be solved. In this paper, we propose a new fusion method that combines depth and color information for foreground segmentation based on an advanced color-based algorithm. First, a background model and a depth model are developed. Then, based on these models, we propose a new updating strategy that can eliminate ghosting and black shadows almost completely. Extensive experiments have been performed to compare the proposed algorithm with other, conventional RGB-D (Red-Green-Blue and Depth) algorithms. The experimental results suggest that our method extracts foregrounds with higher effectiveness and efficiency.

Sudden event recognition: a survey

Event recognition is one of the most active research areas in video surveillance fields. Advancement in event recognition systems mainly aims to provide convenience, safety and an efficient lifestyle for humanity. A precise, accurate and robust approach is necessary to enable event recognition systems to respond to sudden changes in various uncontrolled environments, such as the case of an emergency, physical threat and a fire or bomb alert. The performance of sudden event recognition systems depends heavily on the accuracy of low level processing, like detection, recognition, tracking and machine learning algorithms. This survey aims to detect and characterize a sudden event, which is a subset of an abnormal event in several video surveillance applications. This paper discusses the following in detail: (1) the importance of a sudden event over a general anomalous event; (2) frameworks used in sudden event recognition; (3) the requirements and comparative studies of a sudden event recognition system and (4) various decision-making approaches for sudden event recognition. The advantages and drawbacks of using 3D images from multiple cameras for real-time application are also discussed. The paper concludes with suggestions for future research directions in sudden event recognition.

The Security of IP-Based Video Surveillance Systems

Over the last decade, video surveillance systems have become a part of the Internet of Things (IoT). These IP-based surveillance systems now protect industrial facilities, railways, gas stations, and even one's own home. Unfortunately, like other IoT systems, there are inherent security risks which can lead to significant violations of a user's privacy. In this review, we explore the attack surface of modern surveillance systems and enumerate the various ways they can be compromised with real examples. We also identify the threat agents, their attack goals, attack vectors, and the resulting consequences of successful attacks. Finally, we present current countermeasures and best practices and discuss the threat horizon. The purpose of this review is to provide researchers and engineers with a better understanding of a modern surveillance systems' security, to harden existing systems and develop improved security solutions.

Detecting Malicious False Frame Injection Attacks on Surveillance Systems at the Edge Using Electrical Network Frequency Signals

Over the past few years, the importance of video surveillance in securing national critical infrastructure has significantly increased, with applications including the detection of failures and anomalies. Accompanied by the proliferation of video is the increasing number of attacks against surveillance systems. Among the attacks, False Frame Injection (FFI) attacks that replay video frames from a previous recording to mask the live feed has the highest impact. While many attempts have been made to detect FFI frames using features from the video feeds, video analysis is computationally too intensive to be deployed on-site for real-time false frame detection. In this paper, we investigated the feasibility of FFI attacks on compromised surveillance systems at the edge and propose an effective technique to detect the injected false video and audio frames by monitoring the surveillance feed using the embedded Electrical Network Frequency (ENF) signals. An ENF operates at a nominal frequency of 60 Hz/50 Hz based on its geographical location and maintains a stable value across the entire power grid interconnection with minor fluctuations. For surveillance system video/audio recordings connected to the power grid, the ENF signals are embedded. The time-varying nature of the ENF component was used as a forensic application for authenticating the surveillance feed. The paper highlights the ENF signal collection from a power grid creating a reference database and ENF extraction from the recordings using conventional short-time Fourier Transform and spectrum detection for robust ENF signal analysis in the presence of noise and interference caused in different harmonics. The experimental results demonstrated the effectiveness of ENF signal detection and/or abnormalities for FFI attacks.

Predicting Driver Behavior during the Yellow Interval Using Video Surveillance

At a signalized intersection, drivers must make a stop/go decision at the onset of the yellow signal. Incorrect decisions would lead to red light running (RLR) violations or crashes. This study aims to predict drivers' stop/go decisions and RLR violations during yellow intervals. Traffic data such as vehicle approaching speed, acceleration, distance to the intersection, and occurrence of RLR violations are gathered by a Vehicle Data Collection System (VDCS). An enhanced Gaussian Mixture Model (GMM) is used to extract moving vehicles from target lanes, and the Kalman Filter (KF) algorithm is utilized to acquire vehicle trajectories. The data collected from the VDCS are further analyzed by a sequential logit model, and the relationship between drivers' stop/go decisions and RLR violations is identified. The results indicate that the distance of vehicles to the stop line at the onset of the yellow signal is an important predictor for both drivers' stop/go decisions and RLR violations. In addition, vehicle approaching speed is a contributing factor for stop/go decisions. Furthermore, the accelerations of vehicles after the onset of the yellow signal are positively related to RLR violations. The findings of this study can be used to predict the probability of drivers' RLR violations and improve traffic safety at signalized intersections.

Shadow Detection Based on Regions of Light Sources for Object Extraction in Nighttime Video

Intelligent video surveillance systems detect pre-configured surveillance events through background modeling, foreground and object extraction, object tracking, and event detection. Shadow regions inside video frames sometimes appear as foreground objects, interfere with ensuing processes, and finally degrade the event detection performance of the systems. Conventional studies have mostly used intensity, color, texture, and geometric information to perform shadow detection in daytime video, but these methods lack the capability of removing shadows in nighttime video. In this paper, a novel shadow detection algorithm for nighttime video is proposed; this algorithm partitions each foreground object based on the object’s vertical histogram and screens out shadow objects by validating their orientations heading toward regions of light sources. From the experimental results, it can be seen that the proposed algorithm shows more than 93.8% shadow removal and 89.9% object extraction rates for nighttime video sequences, and the algorithm outperforms conventional shadow removal algorithms designed for daytime videos.

Fall Detection System-Based Posture-Recognition for Indoor Environments

The majority of the senior population lives alone at home. Falls can cause serious injuries, such as fractures or head injuries. These injuries can be an obstacle for a person to move around and normally practice his daily activities. Some of these injuries can lead to a risk of death if not handled urgently. In this paper, we propose a fall detection system for elderly people based on their postures. The postures are recognized from the human silhouette which is an advantage to preserve the privacy of the elderly. The effectiveness of our approach is demonstrated on two well-known datasets for human posture classification and three public datasets for fall detection, using a Support-Vector Machine (SVM) classifier. The experimental results show that our method can not only achieves a high fall detection rate but also a low false detection.

Robust feedback zoom tracking for digital video surveillance

Zoom tracking is an important function in video surveillance, particularly in traffic management and security monitoring. It involves keeping an object of interest in focus during the zoom operation. Zoom tracking is typically achieved by moving the zoom and focus motors in lenses following the so-called “trace curve”, which shows the in-focus motor positions versus the zoom motor positions for a specific object distance. The main task of a zoom tracking approach is to accurately estimate the trace curve for the specified object. Because a proportional integral derivative (PID) controller has historically been considered to be the best controller in the absence of knowledge of the underlying process and its high-quality performance in motor control, in this paper, we propose a novel feedback zoom tracking (FZT) approach based on the geometric trace curve estimation and PID feedback controller. The performance of this approach is compared with existing zoom tracking methods in digital video surveillance. The real-time implementation results obtained on an actual digital video platform indicate that the developed FZT approach not only solves the traditional one-to-many mapping problem without pre-training but also improves the robustness for tracking moving or switching objects which is the key challenge in video surveillance.

Person Re-Identification with RGB-D and RGB-IR Sensors: A Comprehensive Survey

Learning about appearance embedding is of great importance for a variety of different computer-vision applications, which has prompted a surge in person re-identification (Re-ID) papers. The aim of these papers has been to identify an individual over a set of non-overlapping cameras. Despite recent advances in RGB-RGB Re-ID approaches with deep-learning architectures, the approach fails to consistently work well when there are low resolutions in dark conditions. The introduction of different sensors (i.e., RGB-D and infrared (IR)) enables the capture of appearances even in dark conditions. Recently, a lot of research has been dedicated to addressing the issue of finding appearance embedding in dark conditions using different advanced camera sensors. In this paper, we give a comprehensive overview of existing Re-ID approaches that utilize the additional information from different sensor-based methods to address the constraints faced by RGB camera-based person Re-ID systems. Although there are a number of survey papers that consider either the RGB-RGB or Visible-IR scenarios, there are none that consider both RGB-D and RGB-IR. In this paper, we present a detailed taxonomy of the existing approaches along with the existing RGB-D and RGB-IR person Re-ID datasets. Then, we summarize the performance of state-of-the-art methods on several representative RGB-D and RGB-IR datasets. Finally, future directions and current issues are considered for improving the different sensor-based person Re-ID systems.

Real-Time Abnormal Object Detection for Video Surveillance in Smart Cities

With the adaptation of video surveillance in many areas for object detection, monitoring abnormal behavior in several cameras requires constant human tracking for a single camera operative, which is a tedious task. In multiview cameras, accurately detecting different types of guns and knives and classifying them from other video surveillance objects in real-time scenarios is difficult. Most detecting cameras are resource-constrained devices with limited computational capacities. To mitigate this problem, we proposed a resource-constrained lightweight subclass detection method based on a convolutional neural network to classify, locate, and detect different types of guns and knives effectively and efficiently in a real-time environment. In this paper, the detection classifier is a multiclass subclass detection convolutional neural network used to classify object frames into different sub-classes such as abnormal and normal. The achieved mean average precision by the best state-of-the-art framework to detect either a handgun or a knife is 84.21% or 90.20% on a single camera view. After extensive experiments, the best precision obtained by the proposed method for detecting different types of guns and knives was 97.50% on the ImageNet dataset and IMFDB, 90.50% on the open-image dataset, 93% on the Olmos dataset, and 90.7% precision on the multiview cameras. This resource-constrained device has shown a satisfactory result, with a precision score of 85.5% for detection in a multiview camera.

Training Data Extraction and Object Detection in Surveillance Scenario

Police and various security services use video analysis for securing public space, mass events, and when investigating criminal activity. Due to a huge amount of data supplied to surveillance systems, some automatic data processing is a necessity. In one typical scenario, an operator marks an object in an image frame and searches for all occurrences of the object in other frames or even image sequences. This problem is hard in general. Algorithms supporting this scenario must reconcile several seemingly contradicting factors: training and detection speed, detection reliability, and learning from small data sets. In the system proposed here, we use a two-stage detector. The first region proposal stage is based on a Cascade Classifier while the second classification stage is based either on a Support Vector Machines (SVMs) or Convolutional Neural Networks (CNNs). The proposed configuration ensures both speed and detection reliability. In addition to this, an object tracking and background-foreground separation algorithm is used, supported by the GrabCut algorithm and a sample synthesis procedure, in order to collect rich training data for the detector. Experiments show that the system is effective, useful, and applicable to practical surveillance tasks.

Low Cost Efficient Deliverying Video Surveillance Service to Moving Guard for Smart Home

Low-cost video surveillance systems are attractive for Smart Home applications (especially in emerging economies). Those systems use the flexibility of the Internet of Things to operate the video camera only when an intrusion is detected. We are the only ones that focus on the design of protocols based on intelligent agents to communicate the video of an intrusion in real time to the guards by wireless or mobile networks. The goal is to communicate, in real time, the video to the guards who can be moving towards the smart home. However, this communication suffers from sporadic disruptions that difficults the control and drastically reduces user satisfaction and operativity of the system. In a novel way, we have designed a generic software architecture based on design patterns that can be adapted to any hardware in a simple way. The implanted hardware is of very low economic cost; the software frameworks are free. In the experimental tests we have shown that it is possible to communicate to the moving guard, intrusion notifications (by e-mail and by instant messaging), and the first video frames in less than 20 s. In addition, we automatically recovered the frames of video lost in the disruptions in a transparent way to the user, we supported vertical handover processes and we could save energy of the smartphone's battery. However, the most important thing was that the high satisfaction of the people who have used the system.

Behavioral patterns of bats at a wind turbine confirm seasonality of fatality risk

Bat fatalities at wind energy facilities in North America are predominantly comprised of migratory, tree-dependent species, but it is unclear why these bats are at higher risk. Factors influencing bat susceptibility to wind turbines might be revealed by temporal patterns in their behaviors around these dynamic landscape structures. In northern temperate zones, fatalities occur mostly from July through October, but whether this reflects seasonally variable behaviors, passage of migrants, or some combination of factors remains unknown. In this study, we examined video imagery spanning one year in the state of Colorado in the United States, to characterize patterns of seasonal and nightly variability in bat behavior at a wind turbine. We detected bats on 177 of 306 nights representing approximately 3,800 hr of video and > 2,000 discrete bat events. We observed bats approaching the turbine throughout the night across all months during which bats were observed. Two distinct seasonal peaks of bat activity occurred in July and September, representing 30% and 42% increases in discrete bat events from the preceding months June and August, respectively. Bats exhibited behaviors around the turbine that increased in both diversity and duration in July and September. The peaks in bat events were reflected in chasing and turbine approach behaviors. Many of the bat events involved multiple approaches to the turbine, including when bats were displaced through the air by moving blades. The seasonal and nightly patterns we observed were consistent with the possibility that wind turbines invoke investigative behaviors in bats in late summer and autumn coincident with migration and that bats may return and fly close to wind turbines even after experiencing potentially disruptive stimuli like moving blades. Our results point to the need for a deeper understanding of the seasonality, drivers, and characteristics of bat movement across spatial scales.

Normalized Metadata Generation for Human Retrieval Using Multiple Video Surveillance Cameras

Since it is impossible for surveillance personnel to keep monitoring videos from a multiple camera-based surveillance system, an efficient technique is needed to help recognize important situations by retrieving the metadata of an object-of-interest. In a multiple camera-based surveillance system, an object detected in a camera has a different shape in another camera, which is a critical issue of wide-range, real-time surveillance systems. In order to address the problem, this paper presents an object retrieval method by extracting the normalized metadata of an object-of-interest from multiple, heterogeneous cameras. The proposed metadata generation algorithm consists of three steps: (i) generation of a three-dimensional (3D) human model; (ii) human object-based automatic scene calibration; and (iii) metadata generation. More specifically, an appropriately-generated 3D human model provides the foot-to-head direction information that is used as the input of the automatic calibration of each camera. The normalized object information is used to retrieve an object-of-interest in a wide-range, multiple-camera surveillance system in the form of metadata. Experimental results show that the 3D human model matches the ground truth, and automatic calibration-based normalization of metadata enables a successful retrieval and tracking of a human object in the multiple-camera video surveillance system.

Optimal Training Configurations of a CNN-LSTM-Based Tracker for a Fall Frame Detection System

In recent years, there has been an immense amount of research into fall event detection. Generally, a fall event is defined as a situation in which a person unintentionally drops down onto a lower surface. It is crucial to detect the occurrence of fall events as early as possible so that any severe fall consequences can be minimized. Nonetheless, a fall event is a sporadic incidence that occurs seldomly that is falsely detected due to a wide range of fall conditions and situations. Therefore, an automated fall frame detection system, which is referred to as the SmartConvFall is proposed to detect the exact fall frame in a video sequence. It is crucial to know the exact fall frame as it dictates the response time of the system to administer an early treatment to reduce the fall's negative consequences and related injuries. Henceforth, searching for the optimal training configurations is imperative to ensure the main goal of the SmartConvFall is achieved. The proposed SmartConvFall consists of two parts, which are object tracking and instantaneous fall frame detection modules that rely on deep learning representations. The first stage will track the object of interest using a fully convolutional neural network (CNN) tracker. Various training configurations such as optimizer, learning rate, mini-batch size, number of training samples, and region of interest are individually evaluated to determine the best configuration to produce the best tracker model. Meanwhile, the second module goal is to determine the exact instantaneous fall frame by modeling the continuous object trajectories using the Long Short-Term Memory (LSTM) network. Similarly, the LSTM model will undergo various training configurations that cover different types of features selection and the number of stacked layers. The exact instantaneous fall frame is determined using an assumption that a large movement difference with respect to the ground level along the vertical axis can be observed if a fall incident happened. The proposed SmartConvFall is a novel technique as most of the existing methods still relying on detection rather than the tracking module. The SmartConvFall outperforms the state-of-the-art trackers, namely TCNN and MDNET-N trackers, with the highest expected average overlap, robustness, and reliability metrics of 0.1619, 0.6323, and 0.7958, respectively. The SmartConvFall also managed to produce the lowest number of tracking failures with only 43 occasions. Moreover, a three-stack LSTM delivers the lowest mean error with approximately one second delay time in locating the exact instantaneous fall frame. Therefore, the proposed SmartConvFall has demonstrated its potential and suitability to be implemented for a real-time application that could help to avoid any crucial fall consequences such as death and internal bleeding if the early treatment can be administered.

Towards Building a Visual Behaviour Analysis Pipeline for Suicide Detection and Prevention

Understanding human behaviours through video analysis has seen significant research progress in recent years with the advancement of deep learning. This topic is of great importance to the next generation of intelligent visual surveillance systems which are capable of real-time detection and analysis of human behaviours. One important application is to automatically monitor and detect individuals who are in crisis at suicide hotspots to facilitate early intervention and prevention. However, there is still a significant gap between research in human action recognition and visual video processing in general, and their application to monitor hotspots for suicide prevention. While complex backgrounds, non-rigid movements of pedestrians and limitations of surveillance cameras and multi-task requirements for a surveillance system all pose challenges to the development of such systems, a further challenge is the detection of crisis behaviours before a suicide attempt is made, and there is a paucity of datasets in this area due to privacy and confidentiality issues. Most relevant research only applies to detecting suicides such as hangings or jumps from bridges, providing no potential for early prevention. In this research, these problems are addressed by proposing a new modular design for an intelligent visual processing pipeline that is capable of pedestrian detection, tracking, pose estimation and recognition of both normal actions and high risk behavioural cues that are important indicators of a suicide attempt. Specifically, based on the key finding that human body gestures can be used for the detection of social signals that potentially precede a suicide attempt, a new 2D skeleton-based action recognition algorithm is proposed. By using a two-branch network that takes advantage of three types of skeleton-based features extracted from a sequence of frames and a stacked LSTM structure, the model predicts the action label at each time step. It achieved good performance on both the public dataset JHMDB and a smaller private CCTV footage collection on action recognition. Moreover, a logical layer, which uses knowledge from a human coding study to recognise pre-suicide behaviour indicators, has been built on top of the action recognition module to compensate for the small dataset size. It enables complex behaviour patterns to be recognised even from smaller datasets. The whole pipeline has been tested in a real-world application of suicide prevention using simulated footage from a surveillance system installed at a suicide hotspot, and preliminary results confirm its effectiveness at capturing crisis behaviour indicators for early detection and prevention of suicide.

A Dataset of Annotated Omnidirectional Videos for Distancing Applications

Omnidirectional (or 360°) cameras are acquisition devices that, in the next few years, could have a big impact on video surveillance applications, research, and industry, as they can record a spherical view of a whole environment from every perspective. This paper presents two new contributions to the research community: the CVIP360 dataset, an annotated dataset of 360° videos for distancing applications, and a new method to estimate the distances of objects in a scene from a single 360° image. The CVIP360 dataset includes 16 videos acquired outdoors and indoors, annotated by adding information about the pedestrians in the scene (bounding boxes) and the distances to the camera of some points in the 3D world by using markers at fixed and known intervals. The proposed distance estimation algorithm is based on geometry facts regarding the acquisition process of the omnidirectional device, and is uncalibrated in practice: the only required parameter is the camera height. The proposed algorithm was tested on the CVIP360 dataset, and empirical results demonstrate that the estimation error is negligible for distancing applications.

Bus Violence: An Open Benchmark for Video Violence Detection on Public Transport

The automatic detection of violent actions in public places through video analysis is difficult because the employed Artificial Intelligence-based techniques often suffer from generalization problems. Indeed, these algorithms hinge on large quantities of annotated data and usually experience a drastic drop in performance when used in scenarios never seen during the supervised learning phase. In this paper, we introduce and publicly release the Bus Violence benchmark, the first large-scale collection of video clips for violence detection on public transport, where some actors simulated violent actions inside a moving bus in changing conditions, such as the background or light. Moreover, we conduct a performance analysis of several state-of-the-art video violence detectors pre-trained with general violence detection databases on this newly established use case. The achieved moderate performances reveal the difficulties in generalizing from these popular methods, indicating the need to have this new collection of labeled data, beneficial for specializing them in this new scenario.