Advances in Convolution Neural Networks Based Crowd Counting and Density Estimation

Automatically estimating the number of people in unconstrained scenes is a crucial yet challenging task in different real-world applications, including video surveillance, public safety, urban planning, and traffic monitoring. In addition, methods developed to estimate the number of people can be adapted and applied to related tasks in various fields, such as plant counting, vehicle counting, and cell microscopy. Many challenges and problems face crowd counting, including cluttered scenes, extreme occlusions, scale variation, and changes in camera perspective. Therefore, in the past few years, tremendous research efforts have been devoted to crowd counting, and numerous excellent techniques have been proposed. The significant progress in crowd counting methods in recent years is mostly attributed to advances in deep convolution neural networks (CNNs) as well as to public crowd counting datasets. In this work, we review the papers that have been published in the last decade and provide a comprehensive survey of the recent CNNs based crowd counting techniques. We briefly review detection-based, regression-based, and traditional density estimation based approaches. Then, we delve into detail regarding the deep learning based density estimation approaches and recently published datasets. In addition, we discuss the potential applications of crowd counting and in particular its applications using unmanned aerial vehicle (UAV) images.

Accuracy Assessment of UAV-Photogrammetric-Derived Products Using PPK and GCPs in Challenging Terrains: In Search of Optimized Rockfall Mapping

Unmanned aerial photogrammetric surveys are increasingly being used for mapping and studying natural hazards, such as rockfalls. Surveys using unmanned aerial vehicles (UAVs) can be performed in remote, hardly accessible, and dangerous areas, while the photogrammetric-derived products, with high spatial and temporal accuracy, can provide us with detailed information about phenomena under consideration. However, as photogrammetry commonly uses indirect georeferencing through bundle block adjustment (BBA) with ground control points (GCPs), data acquisition in the field is not only time-consuming and labor-intensive, but also extremely dangerous. Therefore, the main goal of this study was to investigate how accurate photogrammetric products can be produced by using BBA without GCPs and auxiliary data, namely using the coordinates X0, Y0 and Z0 of the camera perspective centers computed with PPK (Post-Processing Kinematic). To this end, orthomosaics and digital surface models (DSMs) were produced for three rockfall sites by using images acquired with a DJI Phantom 4 RTK and the two different BBA methods mentioned above (hereafter referred to as BBA_traditional and BBA_PPK). The accuracy of the products, in terms of the Root Mean Square Error (RMSE), was computed by using verification points (VPs). The accuracy of both BBA methods was also assessed. To test the differences between the georeferencing methods, two statistical test were used, namely a paired Student’s t-test, and a non-parametric Wilcoxon signed-rank. The results show that the accuracy of the BBA_PPK is inferior to that of BBA_traditional, with the total RMSE values for the three sites being 0.056, 0.066, and 0.305 m, respectively, compared to 0.019, 0.036 and 0.014 m obtained with BBA_traditional. The accuracies of the BBA methods are reflected in the accuracy of the orthomosaics, whose values for the BBA_PPK are 0.039, 0.043 and 0.157 m, respectively, against 0.029, 0.036 and 0.020 m obtained with the BBA_traditional. Concerning the DSM, those produced with the BBA_PPK method present accuracy values of 0.065, 0.072 and 0.261 m, respectively, against 0.038, 0.060 and 0.030 m obtained with the BBA_traditional. Even though that there are statistically significant differences between the georeferencing methods, one can state that the BBA_PPK presents a viable solution to map dangerous and exposed areas, such as rockfall transit and deposit areas, especially for applications at a regional level.

Video-assisted cardiopulmonary resuscitation: Does the camera perspective matter? A randomized, controlled simulation trial

Background Dispatcher assistance can help to save lives during layperson cardiopulmonary resuscitation during cardiac arrest. The aim of this study was to investigate the influence of different camera positions on the evaluation of cardiopulmonary resuscitation performance during video-assisted cardiopulmonary resuscitation. Methods For this randomized, controlled simulation trial, seven video sequences of cardiopulmonary resuscitation performance were recorded from three different camera positions: side, foot and head position. Video sequences showed either correct cardiopulmonary resuscitation performance or one of the six typical errors: low and high compression rate, superficial and increased compression depth, wrong hand position or incomplete release. Video sequences with different cardiopulmonary resuscitation performances and camera positions were randomly combined such that each evaluator was presented seven individual combinations of cardiopulmonary resuscitation and camera position and evaluated each cardiopulmonary resuscitation performance once. A total of 46 paramedics and 47 emergency physicians evaluated seven video sequences of cardiopulmonary resuscitation performance from different camera positions. The primary hypothesis was that there are differences in accuracy of correct assessment/error recognition depending on camera perspective. Generalized linear multi-level analyses assuming a binomial distribution and a logit link were employed to account for the dependency between each evaluator's seven ratings. Results Of 651 video sequences, cardiopulmonary resuscitation performance was evaluable in 96.8% and correctly evaluated in 74.5% over all camera positions. Cardiopulmonary resuscitation performance was classified correctly from a side perspective in 81.3%, from a foot perspective in 68.8% and from a head perspective in 73.6%, revealing a significant difference in error recognition depending on the camera perspective ( p = .01). Correct cardiopulmonary resuscitation was mistakenly evaluated to be false in 46.2% over all perspectives. Conclusions Participants were able to recognize significantly more mistakes when the camera was located on the opposite side of the cardiopulmonary resuscitation provider. Foot position should be avoided in order to enable the dispatcher the best possible view to evaluating cardiopulmonary resuscitation quality.

GADGETS, PHONES AND DRONES Multi-Perspective Documentary Storytelling

Documentary filmmakers have always been quick to adopt new semi-professional and consumer cameras. However, they have not replaced conventional professional cameras, but added to the vivid variety of documentary style.Today, multi-perspective storytelling in documentary forms is on the rise. On Youtube, Instagram, Snapchat and in TV-Formats various cameras like cellphones, action-cams (GoPros) or drones are commonly used in addition to the classical single-perspective camera. Because the phenomenon is still young, there is very little research and literature on the influence of multi-perspective use of cameras in documentary.Our practice-led, comparative research project ‘Gadgets, Phones and Drones’ investigates the differences of single- and multi-camera storytelling in documentaries and aims to clarify how the use of multi-perspective in documentary is developing. Multi-perspective storytelling is examined by semi-structured interviews with experts in the field and by a practice-based comparative study. In a short documentary about a dog school, we aim to tell the same story in two different ways: We compare the classical shoulder-mounted single-camera-perspective with the multiple camera perspective documenting the very same events. In this process, in the multi-perspective version the dogs as well as their owners and the dog trainer were equipped with cameras and, in addition, the situation was also filmed by a drone.This paper gives insight into questions that arose throughout this artistic research as well as into the discussion of multi-perspective storytelling among practicioners. The mixed method approach will not only add to scientific research, but will also serve as direct feedback for the artistic discussion in current documentary filmmaking.

Visual estimates of blood loss by medical laypeople: Effects of blood loss volume, victim gender, and perspective

A severe hemorrhage can result in death within minutes, before professional first responders have time to arrive. Thus, intervention by bystanders, who may lack medical training, may be necessary to save a victim’s life in situations with bleeding injuries. Proper intervention requires that bystanders accurately assess the severity of the injury and respond appropriately. As many bystanders lack tools and training, they are limited in terms of the information they can use in their evaluative process. In hemorrhage situations, visible blood loss may serve as a dominant cue to action. Therefore, understanding how medically untrained bystanders (i.e., laypeople) perceive hemorrhage is important. The purpose of the current study was to investigate the ability of laypeople to visually assess blood loss and to examine factors that may impact accuracy and the classification of injury severity. A total of 125 laypeople watched 78 short videos each of individuals experiencing a hemorrhage. Victim gender, volume of blood lost, and camera perspective were systematically manipulated in the videos. The results revealed that laypeople overestimated small volumes of blood loss (from 50 to 200 ml), and underestimated larger volumes (from 400 to 1900 ml). Larger volumes of blood loss were associated with larger estimation errors. Further, blood loss was underestimated more for female victims than male victims and their hemorrhages were less likely to be classified as life-threatening. These results have implications for training and intervention design.

Robust Parking Block Segmentation from a Surveillance Camera Perspective

Parking block regions host dangerous behaviors that can be detected from a surveillance camera perspective. However, these regions are often occluded, subject to ground bumpiness or steep slopes, and thus they are hard to segment. Firstly, the paper proposes a pyramidal solution that takes advantage of satellite views of the same scene, based on a deep Convolutional Neural Network (CNN). Training a CNN from the surveillance camera perspective is rather impossible due to the combinatory explosion generated by multiple point-of-views. However, CNNs showed great promise on previous works over satellite images. Secondly, even though there are many datasets for occupancy detection in parking lots, none of them were designed to tackle the parking block segmentation problem directly. Given the lack of a suitable dataset, we also propose APKLOT, a dataset of roughly 7000 polygons for segmenting parking blocks from the satellite perspective and from the camera perspective. Moreover, our method achieves more than 50% intersection over union (IoU) in all the testing sets, that is, at both the satellite view and the camera view.

A Calculation Method for Vehicle Movement Reconstruction from Videos

This paper proposes a new enhanced method based on one-dimensional direct linear transformation for estimating vehicle movement states in video sequences. The proposed method utilizes a contoured structure of target vehicles, and the data collection procedure is found to be relatively stable and effective, providing a better applicability. The movements of vehicles in the video are captured by active calibration regions while the spatial consistency between the vehicle’s driving track and the calibration information are in sync. The vehicle movement states in the verification phase are estimated using the proposed method first, and then the estimated states are compared with the actual movement states recorded in the experimental test. The results show that, in the case of camera perspective of 90 degrees, in all driving states of low speed, high speed, or deceleration, the error between estimated speed and recorded speed is less than 1.5%, the error of accelerations is less than 7%, and the error of distances is less than 2%; similarly, in the case of camera perspective of 30 degrees, the errors of speeds, distances, and accelerations are less than 4%, 5%, and 10%, respectively. It is found that the proposed method is superior to other existing methods.

3D Human Pose Estimation via Explicit Compositional Depth Maps

In this work, we tackle the problem of estimating 3D human pose in camera space from a monocular image. First, we propose to use densely-generated limb depth maps to ease the learning of body joints depth, which are well aligned with image cues. Then, we design a lifting module from 2D pixel coordinates to 3D camera coordinates which explicitly takes the depth values as inputs, and is aligned with camera perspective projection model. We show our method achieves superior performance on large-scale 3D pose datasets Human3.6M and MPI-INF-3DHP, and sets the new state-of-the-art.