Search

Normal behavior is usually more common than abnormal behavior, leading to the problem of category imbalance in the dataset. This can lead to models being biased towards normal behaviors that are easier to identify and less accurate in identifying abnormal behaviors. Meanwhile, the number of people may change in different video frames, and such changes in the number of people can lead to a decrease in the adaptability of most models. To improve the accuracy of abnormal behavior recognition in multi-person scenarios, this paper proposes an intelligent recognition framework based on skeleton pose estimation and person detection, which integrates a pose estimation module, You Only Look Once Human Pose (YH-Pose), and a behavior classification module, Bidirectional Recognition Long Short Term Memory (BR-Lstm). Firstly, the YH-Pose module predicts the 2D positions of human skeletal joint points in the video frames. Secondly, the BR-Lstm module takes as input the skeleton coordinates detached encoding to generate behavioral feature vectors and global feature representations. Finally, a classifier classifies the behavior into normal and abnormal. The proposed framework was experimented on two publicly available datasets: HMDB51 and NTU_RGB+D. The experimental results show that the accuracy of behavioral recognition is 51.3% and 96.7%, respectively, which is better than the mainstream model. The proposed framework was also evaluated with actual surveillance video data. The experimental results showed that the framework could detect five abnormal behaviours on traffic roads: fall , kicking, punching, vomiting and looking down at the mobile phone. The code is avaliable at https://github.com/3083156185/MP-Abr.git. [ABSTRACT FROM AUTHOR]

Person search, which involves identifying target pedestrians in extensive galleries through person detection and re-identification, has experienced significant advancements across various applications. However, it remains a challenging research area due to factors such as appearance changes, lighting variations, background interference, and pedestrian occlusion. This paper proposes an end-to-end person search framework, termed the Global-Response Pedestrian Attention Network (GPAN-PS), designed to tackle these challenges. Specifically, GPAN-PS includes a novel Global Response Pedestrian Attention (GRPA) module that samples pedestrian features using three shared-weight convolutional layers with distinct dilation rates. This enables the network to adaptively select the optimal receptive field through the Squeeze-and-Excitation (SE) module and the Global Response Normalization (GRN) module, enhancing feature stability. Furthermore, we design a GsConvNeXt Head module to bolster feature expressiveness and facilitate inter-channel information interaction. Rather than employing the ConvNeXt (conv5) module as the Box Head for generating refined proposals, our approach employs the GsConvNeXt Head module. This module is also integrated into the Re-ID Head for the extraction of pedestrian features. Both the GRPA and GsConvNeXt Head modules are flexible and adaptable, allowing for seamless integration into other models. Extensive experiments conducted on two benchmark datasets, CUHK-SYSU and PRW, underscore the superior performance of our proposed method. Notably, on the challenging PRW dataset, our approach achieves a mean Average Precision (mAP) of 59.2% and a Top-1 accuracy of 92.2%.

Human-robot collaboration (HRC) is becoming increasingly important in advanced production systems, such as those used in industries and agriculture. This type of collaboration can contribute to productivity increase by reducing physical strain on humans, which can lead to reduced injuries and improved morale. One crucial aspect of HRC is the ability of the robot to follow a specific human operator safely. To address this challenge, a novel methodology is proposed that employs monocular vision and ultra-wideband (UWB) transceivers to determine the relative position of a human target with respect to the robot. UWB transceivers are capable of tracking humans with UWB transceivers but exhibit a significant angular error. To reduce this error, monocular cameras with Deep Learning object detection are used to detect humans. The reduction in angular error is achieved through sensor fusion, combining the outputs of both sensors using a histogram-based filter. This filter projects and intersects the measurements from both sources onto a 2D grid. By combining UWB and monocular vision, a remarkable 66.67% reduction in angular error compared to UWB localization alone is achieved. This approach demonstrates an average processing time of 0.0183s and an average localization error of 0.14 meters when tracking a person walking at an average speed of 0.21 m/s. This novel algorithm holds promise for enabling efficient and safe human-robot collaboration, providing a valuable contribution to the field of robotics. [ABSTRACT FROM AUTHOR]

When the faces of individuals are not clearly identifiable in surveillance videos due to variations in poses, camera viewpoints and occlusions, the appearances of people play a vital role in their identification. Appearance based person re-identification (re-id) summarizes appearances of persons to identify them across multiple non-overlapping camera views. Existing person re-id solutions work on the cropped person images to learn the salient features of a person instead of working on the raw surveillance images, hence these solutions need an independent preliminary phase of preparing cropped person datasets for the training and evaluation purposes. In contrast, the proposed solution works on the raw surveillance images instead of prerequisite of the cropped person dataset and the proposed hierarchical association building among various local parts of the images results in rich person representations for person re-id. In the proposed solution of Smart Surveillance with Simultaneous Person Detection and Re-identification (SSPDR), the complete surveillance video scenes are processed to perform simultaneous person detection and re-identification for all of the persons captured by a surveillance network. We use region proposals based localization scheme for person detection with an increased confidence strategy about the estimation of bounding boxes locations and the person re-identification module learns the hierarchical associations among local salient body parts of a person. Firstly, the proposed re-id module establishes associations among local horizontal strips of two persons, and afterwards it builds associations among local salient sub-patches of already associated pairs of horizontal strips. We address two major re-id challenges i.e. background noise and scale differences using the proposed re-id solution. In context of simultaneous person detection and re-identification, the proposed method is evaluated on publicly available re-id benchmark Person Re-identification in Wild (PRW) as well as on a local surveillance dataset, and attains state-of-the performance. [ABSTRACT FROM AUTHOR]

Person and suspicious activity detection is a major challenge for image-based surveillance systems. However, the accuracy of person detection is affected by several factors, such as the presence of the person, his trajectory, posture, complex background, and object distortion. In this work, we developed a person-focused dataset that includes various behaviors of students in an educational institution, such as cheating, theft of lab equipment, fights, and threatening situations. This dataset ensures consistent and standardized identification annotations for individuals, making it suitable for detection, tracking, and behavioral analysis of individuals. In addition, we have increased the detection accuracy through an improved architecture called YOLOv5 and introduced an efficient method for detecting global and local anomalous behaviors. This method extracts motion features that accurately describe the person's movement, speed, and direction. To evaluate the effectiveness of our proposed approach, we validated it against our proposed database and publicly available benchmark datasets. Our method achieves state-of-the-art detection accuracy, namely 96.12%, with an error rate of 6.68% compared to existing methods. The empirical results show a significant improvement in anomalous activity detection. Our paper concludes with a summary and a discussion of possible future research directions. [ABSTRACT FROM AUTHOR]

Overhead fisheye images can be used for person detection in intelligent monitoring systems. Unlike horizontal images, people in fisheye cameras are generally distributed in any orientation. When the object is rotated, the feature maps from convolutional neural networks have nonlinear variations and lose many orientation features. Transformer can learn the orientation relationship between features. However, a transformer cannot directly extract orientation features and the effectiveness of detecting small objects needs to be improved. In this paper, We propose a novel rotation-equivariant transformer backbone network, which combines group-equivariant convolution with swin transformer to solve these problems. In our proposed model, the rotation-equivariant feature map extracted by group-equivariant convolution contains a large number of orientation features in multiple directions. Aggregates feature in different directions to enhance the communication of orientation features before computing window self-attention. We propose the equivariant-group relation module for evaluating the similarity of the equivariant-group and calculating the aggregation weights. Our network architecture for multi-level receptive field structure can expand the local receptive field to enhance the detection of small objects. The experiments validate that our model achieves state-of-the-art performance on fisheye image datasets MW-R, HABBOF, and CEPDOF. Compared with the swin transformer, the accuracy of our model is improved by 0.3 % , 0.5 % , and 1.3 % , and the accuracy of small object detection in the CEPDOF dataset is improved by 0.73 % . [ABSTRACT FROM AUTHOR]

Abstract Overhead fisheye images can be used for person detection in intelligent monitoring systems. Unlike horizontal images, people in fisheye cameras are generally distributed in any orientation. When the object is rotated, the feature maps from convolutional neural networks have nonlinear variations and lose many orientation features. Transformer can learn the orientation relationship between features. However, a transformer cannot directly extract orientation features and the effectiveness of detecting small objects needs to be improved. In this paper, We propose a novel rotation-equivariant transformer backbone network, which combines group-equivariant convolution with swin transformer to solve these problems. In our proposed model, the rotation-equivariant feature map extracted by group-equivariant convolution contains a large number of orientation features in multiple directions. Aggregates feature in different directions to enhance the communication of orientation features before computing window self-attention. We propose the equivariant-group relation module for evaluating the similarity of the equivariant-group and calculating the aggregation weights. Our network architecture for multi-level receptive field structure can expand the local receptive field to enhance the detection of small objects. The experiments validate that our model achieves state-of-the-art performance on fisheye image datasets MW-R, HABBOF, and CEPDOF. Compared with the swin transformer, the accuracy of our model is improved by 0.3 $$\%$$ % , 0.5 $$\%$$ % , and 1.3 $$\%$$ % , and the accuracy of small object detection in the CEPDOF dataset is improved by 0.73 $$\%$$ % .

Person Search is a task involving pedestrian detection and person re-identification, aiming to retrieve person images matching a given objective attribute from a large-scale image library. The Person Search models need to understand and capture the detailed features and context information of smaller objects in the image more accurately and comprehensively. The current popular Person Search models, whether end-to-end or two-step, are based on anchor boxes. However, due to the limitations of the anchor itself, the model inevitably has some disadvantages, such as unbalance of positive and negative samples and redundant calculation, which will affect the performance of models. To address the problem of fine-grained understanding of target pedestrians in complex scenes and small sizes, this paper proposes a Deformable-Attention-based Anchor-free Person Search model (DAAPS). Fully Convolutional One-Stage (FCOS), as a classic Anchor-free detector, is chosen as the model's infrastructure. The DAAPS model is the first to combine the Anchor-free Person Search model with Deformable Attention Mechanism, applied to guide the model adaptively adjust the perceptual. The Deformable Attention Mechanism is used to help the model focus on the critical information and effectively improve the poor accuracy caused by the absence of anchor boxes. The experiment proves the adaptability of the Attention mechanism to the Anchor-free model. Besides, with an improved ResNeXt+ network frame, the DAAPS model selects the Triplet-based Online Instance Matching (TOIM) Loss function to achieve a more precise end-to-end Person Search task. Simulation experiments demonstrate that the proposed model has higher accuracy and better robustness than most Person Search models, reaching 95.0% of mean Average Precision (mAP) and 95.6% of Top-1 on the CUHK-SYSU dataset, 48.6% of mAP and 84.7% of Top-1 on the Person Re-identification in the Wild (PRW) dataset, respectively. [ABSTRACT FROM AUTHOR]

Person search mainly consists of two submissions, namely Person Detection and Person Re-identification (reID). Existing approaches are primarily based on Faster R-CNN and Convolutional Neural Network (CNN) (e.g., ResNet). While these structures may detect high-quality bounding boxes, they seem to degrade the performance of re-ID. To address this issue, this paper proposes a Dual-Transformer Head Network (DTHN) for end-to-end person search, which contains two independent Transformer heads, a box head for detecting the bounding box and extracting efficient bounding box feature, and a re-ID head for capturing high-quality re-ID features for the re-ID task. Specifically, after the image goes through the ResNet backbone network to extract features, the Region Proposal Network (RPN) proposes possible bounding boxes. The box head then extracts more efficient features within these bounding boxes for detection. Following this, the re-ID head computes the occluded attention of the features in these bounding boxes and distinguishes them from other persons or backgrounds. Extensive experiments on two widely used benchmark datasets, CUHK-SYSU and PRW, achieve state-of-the-art performance levels, 94.9 mAP and 95.3 top-1 scores on the CUHK-SYSU dataset, and 51.6 mAP and 87.6 top-1 scores on the PRW dataset, which demonstrates the advantages of this paper’s approach. The efficiency comparison also shows our method is highly efficient in both time and space. [ABSTRACT FROM AUTHOR]

Fisheye cameras has a large field of view, so it is widely used in scene monitoring, robot navigation, intelligent system, virtual reality panorama, augmented reality panorama and other fields, but person detection under the overhead fisheye camera is still a challenge due to its unique radial geometry and barrel distortion. Generic object detection algorithms do not work well for person detection on panoramic images of the fisheye camera. Recent approaches either use radially aligned bounding boxes to detect persons or improve anchor-based methods to obtain rotated bounding boxes. However, these methods require additional hyperparameters (e.g., anchor boxes) and have low generalization ability. To address this issue, we propose a novel model called Group Equivariant Transformer (GET) which uses the Transformer to directly regress the bounding boxes and rotation angles. GET not need any additional hyperparameters and have generalization ability. In our GET, we uses the Group Equivariant Convolutional Network (GECN) and Multi-Scale Encoder Module (MEM) to extract multi-scale rotated embedding features of overhead fisheye image for Transformer, then we propose an embedding optimization loss to improve the diversity of these features. Finally, we use a Decoder Module (DM) to decode the rotated bounding boxes'information from embedding features. Extensive experiments conducted on three benchmark fisheye camera datasets demonstrate that the proposed method achieves the state of the art. [ABSTRACT FROM AUTHOR]

The coronavirus disease (COVID-19) is primarily disseminated through physical contact. As a precaution, it is recommended that indoor spaces have a limited number of people and at least one meter apart. This study proposes a real-time method for monitoring physical distancing compliance in indoor spaces using computer vision and deep learning techniques. The proposed method utilizes YOLO (You Only Look Once), a popular convolutional neural network-based object detection model, pre-trained on the Microsoft COCO (Common Objects in Context) dataset to detect persons and estimate their physical distance in real time. The effectiveness of the proposed method was assessed using metrics including accuracy rate, frame per second (FPS), and mean average precision (mAP). The results show that the YOLO v3 model had the most remarkable accuracy (87.07%) and mAP (89.91%). On the other hand, the highest fps rate of up to 18.71 was achieved by the YOLO v5s model. The results demonstrate the potential of the proposed method for effectively monitoring physical distancing compliance in indoor spaces, providing valuable insights for future use in other public health scenarios. [ABSTRACT FROM AUTHOR]

Light Detection and Ranging (LiDAR) technology is now becoming the main tool in many applications such as autonomous driving and human–robot collaboration. Point-cloud-based 3D object detection is becoming popular and widely accepted in the industry and everyday life due to its effectiveness for cameras in challenging environments. In this paper, we present a modular approach to detect, track and classify persons using a 3D LiDAR sensor. It combines multiple principles: a robust implementation for object segmentation, a classifier with local geometric descriptors, and a tracking solution. Moreover, we achieve a real-time solution in a low-performance machine by reducing the number of points to be processed by obtaining and predicting regions of interest via movement detection and motion prediction without any previous knowledge of the environment. Furthermore, our prototype is able to successfully detect and track persons consistently even in challenging cases due to limitations on the sensor field of view or extreme pose changes such as crouching, jumping, and stretching. Lastly, the proposed solution is tested and evaluated in multiple real 3D LiDAR sensor recordings taken in an indoor environment. The results show great potential, with particularly high confidence in positive classifications of the human body as compared to state-of-the-art approaches. [ABSTRACT FROM AUTHOR]

Orthogonal Time Frequency Space (OTFS) is a new, promising modulation waveform candidate for the next-generation integrated sensing and communication (ISaC) systems, providing environment-awareness capabilities together with high-speed wireless data communications. This paper presents the original results of OTFS-based person monitoring measurements in the 60 GHz millimeter-wave frequency band under realistic conditions, without the assumption of an integer ratio between the actual delays and Doppler shifts of the reflected components and the corresponding resolution of the OTFS grid. As the main contribution of the paper, we propose the use of the persistent homology technique as a method for processing gathered delay-Doppler responses. We highlight the advantages of the persistent homology approach over the standard constant false alarm rate target detector for selected scenarios. [ABSTRACT FROM AUTHOR]

As seen in the COVID-19 pandemic, one of the most important measures is physical distance in viruses transmitted from person to person. According to the World Health Organization (WHO), it is mandatory to have a limited number of people in indoor spaces. Depending on the size of the indoors, the number of persons that can fit in that area varies. Then, the size of the indoor area should be measured and the maximum number of people should be calculated accordingly. Computers can be used to ensure the correct application of the capacity rule in indoors monitored by cameras. In this study, a method is proposed to measure the size of a prespecified region in the video and count the people there in real time. According to this method: (1) predetermining the borders of a region on the video, (2) identification and counting of people in this specified region, (3) it is aimed to estimate the size of the specified area and to find the maximum number of people it can take. For this purpose, the You Only Look Once (YOLO) object detection model was used. In addition, Microsoft COCO dataset pre-trained weights were used to identify and label persons. YOLO models were tested separately in the proposed method and their performances were analyzed. Mean average precision (mAP), frame per second (fps), and accuracy rate metrics were found for the detection of persons in the specified region. While the YOLO v3 model achieved the highest value in accuracy rate and mAP (both 0.50 and 0.75) metrics, the YOLO v5s model achieved the highest fps rate among non-Tiny models. [ABSTRACT FROM AUTHOR]

Person search aims to locate and identify specific persons from a series of uncropped images, which has achieved a significant impact on many human-related applications, e.g., person activity understanding and person tracking. Person search includes two sub-tasks: person detection and person re-identification. Person detection focuses on finding the commonality of all identities, while person re-identification focuses on finding the differences among different identities. To mitigate the impact of the different purposes of these two sub-tasks on a person search model, we split the ResNet50 network according to the sub-task, and propose a feature diversity person search (FDPS) framework based on the rank constraint of the cross-class matrix. We first construct a model called the split-baseline (S-bsl), and then introduce the deformable convolution to locate the entire person area. More importantly, a rank perception optimization (RPO) loss is proposed in the FDPS framework to enhance the discrimination and diversity of inter-class features. Experimental results on PRW and CUHK-SYSU datasets demonstrate the effectiveness of the proposed method. [ABSTRACT FROM AUTHOR]

As security is one of the basic human needs, we need security systems that can prevent crimes from happening. In general, surveillance videos are used to observe the environment and human behavior in a given location. However, surveillance videos can only be used to record images or videos, without additional information. Therefore, more advanced cameras are needed to obtain other additional information such as the position and movement of people. This research extracted this information from surveillance video footage using a person tracking, detection, and identification algorithm. The framework for these is based on deep learning algorithms, a popular branch of artificial intelligence. In the field of video surveillance, person tracking is considered a challenging task. Many computer vision, machine learning, and deep learning techniques have been developed in recent years. The majority of these techniques are based on frontal view images or video sequences. In this work, we will compare some previous work related to the same topic. [ABSTRACT FROM AUTHOR]

Artificial intelligence (AI), combined with the Internet of Things (IoT), plays a beneficial role in various fields, including intelligent surveillance applications. With IoT and 5G advancement, intelligent sensors, and devices in the surveillance environment collect large amounts of data in the form of videos and images. These collected data require intelligent information processing solutions, help analyze the recorded videos and images to detect and identify various objects in the scene, particularly humans. In this study, an automated human detection system is presented for a complex industrial environment, in which people are monitored/detected from a top view perspective. A top view is usually preferred because it can provide sufficient coverage and enough visibility of a scene. This study demonstrates the applications, efficiency, and effectiveness of deep learning architectures, that is, Faster Region Convolutional Neural Network (Faster R‐CNN), Single Shot MultiBox Detector (SSD), and You Only Look Once (YOLOv3), with transfer learning. Experimental results reveal that with additional training and transfer learning, the performance of all detection architectures is significantly improved. The detection results are also compared using the same data set. The deep learning architectures achieve promising results with maximum true‐positive rate of 93%, 94%, and 94% for Faster‐RCNN, SSD, and YOLOv3, respectively. Furthermore, a detailed study is performed on output results that highlight challenges and probable future trends. [ABSTRACT FROM AUTHOR]

The social distancing regulations introduced to slow down the spread of COVID-19 virus directly affect a basic form of non-verbal communication, and there may be longer term impacts on human behavior and culture that remain to be analyzed in proxemics studies. To obtain quantitative results for such studies, large media and/or personal photo collections must be analyzed. Several social distance monitoring methods have been proposed for safety purposes, but they are not directly applicable to general photo collections with large variations in the imaging setup. In such studies, the interest shifts from safety to analyzing subtle differences in social distances. Currently, there is no suitable benchmark for developing such algorithms. Collecting images with measured ground-truth pair-wise distances using different camera settings is cumbersome. Moreover, performance evaluation for these algorithms is not straightforward, and there is no widely accepted evaluation protocol. In this paper, we provide an image dataset with measured pair-wise social distances under different camera positions and settings. We suggest a performance evaluation protocol and provide a benchmark to easily evaluate such algorithms. We also propose an automatic social distance estimation method that can be applied on general photo collections. Our method is a hybrid method that combines deep learning-based object detection and human pose estimation with projective geometry. The method can be applied on uncalibrated single images with known focal length and sensor size. The results on our benchmark are encouraging with 91% human detection rate and only 38.24% average relative distance estimation error among the detected people.

The article analyzes the possibility and rationality of using proctoring technology in remote monitoring of the progress of university students as a tool for identifying a student. Proctoring technology includes face recognition technology. Face recognition belongs to the field of artificial intelligence and biometric recognition. It is a very successful application of image analysis and understanding. To implement the task of determining a person's face in a video stream, the Python programming language was used with the OpenCV code. Mathematical models of face recognition are also described. These mathematical models are processed during data generation, face analysis and image classification. We considered methods that allow the processes of data generation, image analysis and image classification. We have presented algorithms for solving computer vision problems. We placed 400 photographs of 40 students on the base. The photographs were taken at different angles and used different lighting conditions; there were also interferences such as the presence of a beard, mustache, glasses, hats, etc. When analyzing certain cases of errors, it can be concluded that accuracy decreases primarily due to images with noise and poor lighting quality. [ABSTRACT FROM AUTHOR]

Person detection has attracted great attention in the computer vision area and is an imperative element in human-centric computer vision. Although the predictive performances of person detection networks have been improved dramatically, they are vulnerable to adversarial patch attacks. Changing the pixels in a restricted region can easily fool the person detection network in safety-critical applications such as autonomous driving and security systems. Despite the necessity of countering adversarial patch attacks, very few efforts have been dedicated to defending person detection against adversarial patch attack. In this paper, we propose a novel defense strategy that defends against an adversarial patch attack by optimizing a defensive frame for person detection. The defensive frame alleviates the effect of the adversarial patch while maintaining person detection performance with clean person. The proposed defensive frame in the person detection is generated with a competitive learning algorithm which makes an iterative competition between detection threatening module and detection shielding module in person detection. Comprehensive experimental results demonstrate that the proposed method effectively defends person detection against adversarial patch attacks. [ABSTRACT FROM AUTHOR]

Light Detection and Ranging (LiDAR) technology is now becoming the main tool in many applications such as autonomous driving and human–robot collaboration. Point-cloud-based 3D object detection is becoming popular and widely accepted in the industry and everyday life due to its effectiveness for cameras in challenging environments. In this paper, we present a modular approach to detect, track and classify persons using a 3D LiDAR sensor. It combines multiple principles: a robust implementation for object segmentation, a classifier with local geometric descriptors, and a tracking solution. Moreover, we achieve a real-time solution in a low-performance machine by reducing the number of points to be processed by obtaining and predicting regions of interest via movement detection and motion prediction without any previous knowledge of the environment. Furthermore, our prototype is able to successfully detect and track persons consistently even in challenging cases due to limitations on the sensor field of view or extreme pose changes such as crouching, jumping, and stretching. Lastly, the proposed solution is tested and evaluated in multiple real 3D LiDAR sensor recordings taken in an indoor environment. The results show great potential, with particularly high confidence in positive classifications of the human body as compared to state-of-the-art approaches.

We present TIMo (Time-of-flight Indoor Monitoring), a dataset for video-based monitoring of indoor spaces captured using a time-of-flight (ToF) camera. The resulting depth videos feature people performing a set of different predefined actions, for which we provide detailed annotations. Person detection for people counting and anomaly detection are the two targeted applications. Most existing surveillance video datasets provide either grayscale or RGB videos. Depth information, on the other hand, is still a rarity in this class of datasets in spite of being popular and much more common in other research fields within computer vision. Our dataset addresses this gap in the landscape of surveillance video datasets. The recordings took place at two different locations with the ToF camera set up either in a top-down or a tilted perspective on the scene. Moreover, we provide experimental evaluation results from baseline algorithms. [ABSTRACT FROM AUTHOR]

Due to a growing number of people who carry out various adrenaline activities or adventure tourism and stay in the mountains and other inaccessible places, there is an increasing need to organize a search and rescue operation (SAR) to provide assistance and health care to the injured. The goal of SAR operation is to search the largest area of the territory in the shortest time possible and find a lost or injured person. Today, drones (UAVs or drones) are increasingly involved in search operations, as they can capture a large, controlled area in a short amount of time. However, a detailed examination of a large amount of recorded material remains a problem. Even for an expert, it is not easy to find searched people who are relatively small considering the area where they are, often sheltered by vegetation or merged with the ground and in unusual positions due to falls, injuries, or exhaustion. Therefore, the automatic detection of persons and objects in images/videos taken by drones in these operations is very significant. In this paper, the reliability of existing state-of-the-art detectors such as Faster R-CNN, YOLOv4, RetinaNet, and Cascade R-CNN on a VisDrone benchmark and custom-made dataset SARD build to simulate rescue scenes was investigated. After training the models on selected datasets, detection results were compared. Because of the high speed and accuracy and the small number of false detections, the YOLOv4 detector was chosen for further examination. YOLOv4 model results related to different network sizes, different detection accuracies, and transfer learning settings were analyzed. The model robustness to weather conditions and motion blur were also investigated. The paper proposes a model that can be used in SAR operations because of the excellent results in detecting people in search and rescue scenarios.