RETRACTED: Creating an original technique for identifying objects in video streams

The development of mathematical models for detecting moving objects in real-time video streams is a significant task, especially in light of the growing demands for security and process automation. This paper proposes a detection system based on the Faster Region-Based Convolutional Neural Network (Faster R-CNN) model, utilizing various backbone architectures, including Residual Network-50 (ResNet-50), Feature Pyramid Network (FPN), MobileNet Version 3 Large, and Efficient Network-B0 (EfficientNetB0) with a Self-Attention mechanism. The model extracts object features by combining deep neural networks and scaling mechanisms, enabling adaptation to various operating conditions. Experimental results demonstrate that the proposed approach achieves high accuracy, with the MobileNet Version 3 Large model achieving 95.70% accuracy and the Residual Network-50 model reaching 100% accuracy within the first three training cycles (epochs). The Faster Region-Based Convolutional Neural Network model using Efficient Network-B0 with the Self-Attention mechanism reaches 100% accuracy by the third training cycle, maintaining consistent performance throughout the remaining training cycles. The model with the Residual Network-50 and Feature Pyramid Network backbone architecture shows an average loss reduction from 0.0922 in the first training cycle to 0.0102 by the 15th training cycle, confirming its high stability and performance. Future research could focus on optimizing the proposed approach to handle even more complex and dynamic video sequences and improving data processing to reduce computational costs and increase processing speed.

Recent research in computer vision is increasingly focusing on developing systems to understand people’s appearance, movements and activities, provide advanced interfaces for interacting with people, create human models. For any of these systems to work, they need methods to identify people from a particular input image or video. Today, real-time object detection and sizing of objects is an important issue in many areas of the industry. This is a vital issue of computer vision problems. With Covid-19's healing process, it will be very important to maintain social distance. In this research and development, it is aimed to maintain social distance with proposed big data architecture. This article provides an advanced technique to detect objects in video streams in real time and calculate their distance. The system composed research and developments to perform a stream from the camera, such as video stream, distance and object detection model, incoming data stream, data stream collection and report generation. The video stream from the camera is processed with GStream. The frames from the video stream are taken by OpenCV, YOLOV3 is trained by distance and object detection model and developed by Python. Video streaming data trained with Kubeflow is published with Apache Kafka and Apache Spark. It uses HDFS used to store published data. It is used to query and analyze data in Hive, Impala, Hbase HDFS. After that Analytical reports are created. E-mail notifications can be created according to the data in the database using by Apache Oozie. Through the proposed real time big data architecture, people can be safe in closed areas.

This article proposes a new approach which helps to prevent the formation of self-crossings and loops in the parametric active contour model while tracking moving objects in video streams. The presented solutions mean that the process of tracking is stable in all subsequent video sequences. On the other hand, self-crossings of nodes and the resultant contour loops happen very often in the basic model and ruin the tracking process. The Gaussian filter was used to eliminate noise from video streams.

The MPEG-4 video coding standard consists of object based coding schemes for multimedia and enables content based functionalities. Video objects in still pictures or video sequences should be first identified before the encoding process starts. An algorithm based on information fusion, which can be used for the extraction of foreground objects in video streams in real time is proposed in this paper. The method efficiently integrates image and motion information of video streams. The thresholding technique operated in the HSV space provides a better use of the color information than that in the traditional RGB space. Using enhanced boundary extracted from the motion region for contour adaptation offers an original method to refine the contour.

Tracking of moving objects in video streams with considering different dynamic backgrounds is a challenging problem in a real time dynamic environment. The proposed approach uses an observation model based on optical flow information used to know the displacement of the objects present in the scene. After finding the moving regions in the initial frame, we are applying active contour model (ACM) to track the moving objects in the further frames dynamically. These models have been used as a natural means of incorporating flow information into the tracking. The formulation of the active contour model by incorporating an additional force driven optical flow field improves the tracking speed. This algorithm efficiently works to track for low contrast videos like aerial videos.

Multimedia information is ubiquitous and essential in many applications from homeland security to medicine and bioinformatics. As evidenced by the success of the previous editions of MDM/KDD, there is an increasing need in new techniques and tools that can detect and discover patterns, in multimedia data, that can lead to new knowledge. For example, tools are needed for discovering relationships between objects or segments within images, classifying images based on their content, extracting patterns in sound, categorizing speech and music, and recognizing and tracking objects in video streams. There is also an increasing interest in the real-time analysis of multimedia data generated by distributed sensory applications and ambient intelligence environments.MDM is a leading venue where researchers, both from the academia and industry, can exchange and compare both relatively mature and green house theories, methodologies, algorithms and frameworks for multimedia data mining. To address this aim, the workshop brings together experts in the analysis of digital media content, multimedia databases, multimedia information retrieval, and domain experts from different applied disciplines with potential in multimedia data mining and knowledge discovery. The previous seven workshops have been held in conjunction with KDD 2000 (Boston, MA), KDD2001 (San Francisco, CA), KDD 2002 (Edmonton, Canada), KDD 2003 (Washington, DC), KDD 2004 (Seattle, WA), KDD 2005 (Chicago, IL), and KDD 2006 (Philadelphia, PA) respectively. Like the previous editions, MDM 2007 aimed at facilitating cross-disciplinary exchange of ideas.MDM 2007 contains six regular papers focusing on various aspects of multimedia data mining, including content-supported learning of user preferences, integration of multiple visual learners for media mining, salient-point based image object detection, the use of relevance feedback for video event mining, association rule mining for infrequent database items, and a novel algebra for supporting ranked data processing in multimedia mining applications. In addition, we have invited two leading researchers to present their views on major challenges faced by the multimedia data mining community: The invited talk by Dr. Belle Tseng from NEC Labs, America, focuses on the problem of temporal graph mining and its applications to blog analysis. Prof. Keogh from University of California, Riverside, on the other hand, highlights the challenges faced by the data mining community due to frequently non-reproducible results published at the best data mining conferences.We are thankful to the authors who have submitted papers to the workshop, to our invited speakers, to the PC members and reviewers who all have done a fabulous job, to KDD 2007 organizers (especially to Qiang Yang from University of Science and Technology, Hong Kong for his support to the workshop and to Scott Gaffney for his help with the proceedings).

Background Estimation is a common computer vision task, used for segmenting moving objects in video streams. This can be useful as a pre-processing step, isolating regions of interest for more complicated algorithms performing detection, recognition, and identification tasks, in order to reduce overall computation time. This is especially important in the context of embedded systems like smart cameras, which may need to process images with constrained computational resources. This work focuses on accelerating SuperBE, a superpixel-based background estimation algorithm that was designed for simplicity and reducing computational complexity while maintaining state-of-the-art levels of accuracy. We explore both software and hardware acceleration opportunities, converting the original algorithm into a greyscale, integer-only version, and using Hardware/Software Co-design to develop hardware acceleration components on FPGA fabric that assist a software processor. We achieved a 4.4× speed improvement with the software optimisations alone, and a 2× speed improvement with the hardware optimisations alone. When combined, these led to a 9× speed improvement on a Cyclone V System-on-Chip, delivering almost 38 fps on 320 × 240 resolution images.

Computer vision is a relevant technology, that allows the machine, relying on external data from sensors, to independently make decisions on how to react to external factors. One of the possible problems that computer vision could solve is the automatic detection and tracking of point objects. For this purpose, in practice, a video surveillance system is often used, which allows to obtain a video stream of the viewed site. Digital image processing methods allow reliable and precision detection and tracking of various objects. Since the point object is maneuverable, the detection algorithm must be fast and robust to avoid false detections, which makes such algorithms complicated. The work presents an overview of known methods of detecting and tracking objects. The advantages and disadvantages of the methods based on the principle of "detect before track" and "track before detect" are presented. Methods that are based on "detect before track" principle detect the object in each frame, therefore do not require additional accumulation of information about the parameters of the searched object and after that the obtained results are transferred to the tracking algorithm. Methods based on "track before detect" principle first accumulate some frames to determine the trajectory of the object, after that a decision is made whether object detected or not. Such algorithms require a certain amount of data to process, which leads to a delay in obtaining results. The methods of tracking objects are considered. Such methods are based on the processing of coordinate information about the movement of objects. One of the crucial tasks for tracking is the filtering of object motion parameters. To estimate the condition and control a dynamic system with a random structure, it is advisable to apply the apparatus of mixed Markov processes in the discrete time. The following algorithms of trajectory filtering are considered: an autonomous multi-model algorithm, a generalized pseudo-Bayesian algorithm of the first order, a generalized pseudo-Bayesian algorithm of the second order, a multi-model algorithm with inter-model interaction.

Geometric variations of spatial and temporal features of objects in video streams cause great difficulty in video salient object detection (VSOD) tasks. Most existing deep-learning methods utilize fixed-sized kernels, which limits the receptive field (RF) to extract the local and global features and fails to understand the visual semantics of the deformed objects’ foreground and background. Moreover, due to their complex architectures, these methods need more computational resources, which limits their deployment in real-world scenarios. To address the aforementioned challenges and to make a balance between performance and computational complexity, a deformable separable fusion network (DSFNet) is proposed, which extracts the geometric spatiotemporal variations at multiscale features dynamically without compromising the network’s complexity. A Swarm-Enhanced Adam (SEAdam) optimizer has been proposed to adaptively balance the exploration and exploitation of gradients locally and globally and improve the convergence speed. This is the first work that extracts the multiscale geometric local and global context-based visual information. With the help of extensive experimentation on six benchmark highly challenging datasets, we show that the proposed model outperforms state-of-the-art models in terms of the number of parameters, floating-point operations (FLOPs), and latency.