Privacy-Preserving Object Detection with Veracruz

Object detection and recognition is an integral part of Computer Vision. It has a multitude of applications ranging from character recognition to video analysis. Object detection now play a crucial role in industries like security, video, medical, sports, and many more. With the latest research, rapid development in deep learning, and computational image understanding object detection will be more prevalent. In the future, we will be able to implement object detection in much more mundane workings of our life. Our review paper will focus on the existing techniques that are present in the community and how each technique is different from the other techniques. We will start with an introduction about object detection as a tool and technique. This will be followed by a brief background discussing the most fundamental steps involved and the basic architecture of object detection. Then, we focus on several existing techniques and perform comparative analyses of these techniques to draw meaningful conclusions. These techniques include two-stage models such as R-CNN, Fast R-CNN, and Faster R-CNN and single-stage models YOLOv1, YOLOv2, YOLOv3, YOLOv4, and SSD. Finally, we will conclude by presenting some directions that provide a scope for further analyses along that direction and hope to suggest different techniques that can be implemented in different sectors for optimal results.

Cloud computing infrastructures have become the de-facto platform for data driven machine learning applications. However, these centralized models of computing are unqualified for dispersed high-volume real-time edge data intensive applications such as real time object detection, where video streams may be captured at multiple geographical locations. While many recent advancements in object detection have been made using Convolutional Neural Networks, these performance improvements only focus on a single, contiguous object detection model. In this paper, we propose a distributed Edge-Cloud R-CNN pipeline. By splitting the object detection pipeline into components and dynamically distributing these components in the cloud, we can achieve optimal performance to enable real time object detection. As a proof of concept, we evaluate the performance of the proposed system on a distributed computing platform including cloud servers and edge-embedded devices for real-time object detection on live video streams.

The article explores modern neural network architectures for the automatic detection and recognition of marine surface objects and obstacles of given classes throughout the full image area, applicable for execution in real or near real time on an optoelectronic vision system to au-tomate and improve the safety of civil marine navigation. A formal statement of the problem of automatic detection of objects on images is given. The state-of-the-art algorithms for detecting objects in images based on use of artificial convolutional neural networks were reviewed, their comparison was made and a reasonable choice was made in favor of the most efficient neuralnetwork architecture in terms of computational complexity to recognition accuracy. The subject area is studied, as well as publicly available databases of surface objects suitable for use in the training of algorithms using artificial neural networks. The article concluded that there is insuffi-cient labeled data for training neural network algorithms, as a result of which the authors inde-pendently collected research images and video sequences, prepared and labeled the collected data containing surface marine objects and other obstacles that represent a navigation hazard for ships. Based on the selected neural network architecture, a new neural network algorithm for automatic full-frame detection and recognition of surface objects was developed, and an artificial neural network was trained using the prepared database of images of typical objects. The resulting algorithm was tested by the authors on a validation data set, the quality of its work was estimated using various metrics, and the algorithm’s performance was measured. Conclusions are made about the necessity to expand the collected database of images of typical marine objects, further steps are proposed to improve the accuracy of the developed software and algorithmic complex and its implementation to be used in a marine optoelectronic machine vision system for automa-tion and improving the safety of civil navigation.

Multispectral systems for detection of concealed dangerous objects are becoming more popular because of their higher effectiveness compared to mono-spectral systems. So far, the problem of detecting objects hidden under clothing was considered only in the case of airports but it is becoming more important for public places like metro stations, and government buildings. Exploration of new spectral bands as well as development of technology result in introduction of new solutions – both mono and multispectral. It has been proved that objects hidden under clothing can be detected and visualized using terahertz (THz) cameras. However, passive THz cameras still offer too low image resolution for objects recognition. Limited range is another issue of passive imagers. On the other hand new infrared cameras offer sufficient parameters to detect objects covered with fabrics in some conditions, as well as high image quality and big pixel resolutions. The purpose of the studies is to investigate and compare the possibilities of using passive cameras operating in long wavelength infrared (LWIR) and THz spectral ranges for detection of concealed objects. For the purpose of investigations, commercial imagers operating in 6.5-11.7 μm and 250GHz (1.25mm) were used. In the article, we present the measurement setup and the results of measurements in various operating conditions. Theoretical studies of both spectral bands focused on detection of objects with passive imagers are also presented.

The Web contains a lot of information but computers can't do very much with it. We have access to and a need for a lot of information (thanks to pervasive information devices such as mobile phones, PDAs, smart devices, TV, radio...) but we don't have time to find and make associations between information every time we want to. Given the need for quick, intuitive access to information, multimedia presentation is an effective form of communication - different media act as a better means to communicate dependant on the context, and synchronization (organisation of media in space and in time) can express non- verbally relationships between the content (e.g. linear presentation to represent a timeline of events). Appropriately delivering content in the form of synchronized multimedia is a complex task which needs to meet the requirements of the user, device, and actual context. This proves to be too complex to be fulfilled by manually authoring for all possibilities. This has lead to efforts in developing systems for (semi-)automatic multimedia generation. Knowledge representation techniques developed in earlier AI work are being applied to the World Wide Web. Through standards such as RDF and OWL, the so-called "Semantic Web" offers the potential of a large scale distributed knowledge repository built upon the existing Web infrastructure. This makes powerful amounts of knowledge available to Web-based applications as the basis for a new paradigm of Web-oriented services which are intelligent, dynamic, flexible and user-centred. Such services could find the information we need and present it to us in an effective manner. The current situation of Web content discovery and delivery displays significant disorder and inflexibility: natural language is problematically ambiguous and Web content is only adaptive to a limited extent. Using the Semantic Web, intelligent Web clients will be able to deliver the correct information every time by unambiguously understanding the meaning of queries and reasoning on them (through the knowledge available to them on the Web) to come to an answer. Likewise, Web resources will be described with metadata that supports their selection, adaptation and presentation to the needs of the client. Such a paradigm shift in Web-based systems (from content to knowledge-centred) requires a similar paradigm shift in system models, architectures, implementations and operations. This is equally the case for multimedia presentation systems. We propose a framework for a new type of multimedia presentation system in which the multimedia generation process is tightly integrated with the distributed knowledge of the Semantic Web. Unlike previous work in which knowledge is narrowly defined and used by the system, this approach is built with the scale and dynamics of the Web in mind. We name this framework SWeMPs - a Semantic Web-enabled Multimedia Presentation System. The key results of our research are the conceptual model for a multimedia presentation system and the generic rule base which interacts with that model (through querying, reasoning and deriving new knowledge) to realise individual multimedia generation tasks. These form the basis for an implementation of a multimedia presentation system which is fundamentally different from existing multimedia presentation systems in that it operates at the logical or conceptual level rather than the data or syntactic level. This reflects the reality that what users seek is information (knowledge about concepts) and effective communication of that information (respecting the relationships between concepts). It leverages the potential of the Semantic Web as a distributed, large scale, accessible repository of knowledge about the world of the user just as systems today use the Web as a repository of content. The implementation of this proposal serves as a 'proof of concept' and as a basis for the evaluation of the new approach through use cases.

The state-of-the-art object detection method is complicated with various modules such as backbone, RPN, feature fusion neck and RCNN head, where each module may have different designs and structures. How to leverage the computational cost and accuracy trade-off for the structural combination as well as the modular selection of multiple modules? Neural architecture search (NAS) has shown great potential in finding an optimal solution. Existing NAS works for object detection only focus on searching better design of a single module such as backbone or feature fusion neck, while neglecting the balance of the whole system. In this paper, we present a two-stage coarse-to-fine searching strategy named Structural-to-Modular NAS (SM-NAS) for searching a GPU-friendly design of both an efficient combination of modules and better modular-level architecture for object detection. Specifically, Structural-level searching stage first aims to find an efficient combination of different modules; Modular-level searching stage then evolves each specific module and pushes the Pareto front forward to a faster task-specific network. We consider a multi-objective search where the search space covers many popular designs of detection methods. We directly search a detection backbone without pre-trained models or any proxy task by exploring a fast training from scratch strategy. The resulting architectures dominate state-of-the-art object detection systems in both inference time and accuracy and demonstrate the effectiveness on multiple detection datasets, e.g. halving the inference time with additional 1% mAP improvement compared to FPN and reaching 46% mAP with the similar inference time of MaskRCNN.

Object detection in embedded systems is important for many contemporary applications that involve vision and scene analysis. In this paper, we propose a novel architecture for object detection implemented in FPGA, based on the Stripe Memory Engine (SME), and point out shortcomings of existing architectures. SME processes a stream of image data so that it stores a narrow stripe of the input image and its scaled versions and uses a detector unit which is efficiently pipelined across multiple image positions within the SME. We show how to process images with up to 4K resolution at high frame rates using cascades of SMEs. As a detector algorithm, the SMEs use boosted soft cascade with simple image features that require only pixel comparisons and look-up tables; therefore, they are well suitable for hardware implemenation. We describe the components of our architecture and compare it to several published works in several configurations. As an example, we implemented face detection and license plate detection applications that work with HD images (1280 ×720 pixels) running at over 60 frames/s on Xilinx Zynq platform. We analyzed their power consumption, evaluated the accuracy of our detectors, and compared them to Haar Cascades from OpenCV that are often used by other authors. We show that our detectors offer better accuracy as well as performance at lower power consumption.

The demand for sophisticated security solutions has surged due to evolving security threats. This research introduces an advanced intelligent surveillance system that leverages deep learning and computer vision to automate threat detection and provide immediate alerts. Utilizing the YOLO (You Only Look Once) architecture for real-time object detection, combined with OpenCV for video processing and TensorFlow for model optimization, the system offers automated threat classification, dual-alert mechanisms (local sound alarms and email notifications), and customizable detection settings. A web-based interface developed using Flask facilitates user interaction, while MySQL manages detection logs, user preferences, and system analytics. The system demonstrates significant improvements in response time, detection accuracy, and operational efficiency over traditional CCTV systems, making it ideal for residential, commercial, and industrial security applications. Keywords: Intelligent surveillance, deep learning, computer vision, threat detection, YOLO, automated security, real-time monitoring, object detection.

One of the negative consequences of warfare is the contamination of territories with explosive ordnance. Due to the significant contamination of the territories of dozens of countries with explosive objects, the issue of guaranteed demining has become a global problem. Ukraine is no exception, as it has become the most mined country in the world as a result of Russia's war against Ukraine. The most difficult task of demining is to find and detect explosive objects in the environments that cover them. In these conditions, methods and ways to guarantee the detection of explosive objects are being searched for. One of these ways is the development and implementation of unmanned demining systems. The paper presents the results of theoretical research on improving the mathematical model of detecting explosive objects. The main stages of the demining process are considered. An analysis of research related to demining processes, in particular the search and detection of explosive objects, is carried out. Based on the results of the analysis of existing research in this subject area, ways to improve mathematical models of detection of explosive objects are identified - taking into account the reliability of search elements and a mobile platform. An approach to improving the mathematical models for determining the probability of detecting explosive objects is proposed on the example of search elements with a radio wave method of detecting explosive objects. It is proposed to use this approach to determine the probability of detecting explosive objects of remotely controlled (robotic) demining means (complexes), the search elements of which work with the use of other electromagnetic methods of search and detection of explosive objects.

Object Detection methods are an important subject in the implementations of Artificial Intelligent systems. Many attempts to build up a real-time object detection hardware/software in an SoC are presented in recent research. However, due to the demanding in both memory bandwidth and parallel computing resources, only few designs can fulfill the real-time requirement, which is important for applications or payloads such as Drones, UAVs, and autonomous vehicles. In this paper, issues in design of an SoC based object-detection will be discussed, and an on-going hardware design base on YOLO algorithm and ARC Platform will be presented. With the optimization in both numbers of Processing Element and Memory Bandwidth, An 176.3 GOPs CNN accelerator with 30 fps performance at 400MHz is presented. In addition to the Object-Detection engine, a ZCA image preprocessor and NMS postprocessing are also proposed to simplify the corresponding CNN model and enhance the real-time performance of Object-Detection. The emulation results and demonstration videos in the ARC platform will be presented. The post-layout simulation of current designverified the targeting real-time performance in a 28nm CMOS technology.

The analysis of video footage regarding the identification of persons at defined locations or the detection of complex activities is still a challenging process. Nowadays, various (semi-)automated systems can be used to overcome different parts of these challenges. Object detection and their classification reach even higher detection rates when making use of the latest cutting-edge convolutional neural network frameworks. Integrated into a scalable infrastructure as a service data base system, we employ the combination of such networks by using the Detectron framework within Docker containers with case-specific engineered tracking and motion pattern heuristics in order to detect several activities with comparatively low and distributed computing efforts and reasonable results.

Image restoration algorithms such as super resolution (SR) are indispensable pre-processing modules for object detection in low quality images. Most of these algorithms assume the degradation is fixed and known a priori. However, in practical, either the real degradation or optimal up-sampling ratio rate is unknown or differs from assumption, leading to a deteriorating performance for both the pre-processing module and the consequent high-level task such as object detection. Here, we propose a novel self-supervised framework to detect objects in degraded low resolution images. We utilizes the downsampling degradation as a kind of transformation for self-supervised signals to explore the equivariant representation against various resolutions and other degradation conditions. The Auto Encoding Resolution in Self-supervision (AERIS) framework could further take the advantage of advanced SR architectures with an arbitrary resolution restoring decoder to reconstruct the original correspondence from the degraded input image. Both the representation learning and object detection are optimized jointly in an end-to-end training fashion. The generic AERIS framework could be implemented on various mainstream object detection architectures with different backbones. The extensive experiments show that our methods has achieved superior performance compared with existing methods when facing variant degradation situations. Code is available at this link https://github.com/cuiziteng/ECCV_AERIS.KeywordsSelf-supervised learningComputational photographyObject detection

As one of the hottest fields of computer vision, object detection is employed in many useful applications, including face recognition, autonomous cars and human behaviour analysis. CNN's have played a vital role in the progression of computer vision, especially object detection. In this paper, a review is provided on object detection methods which utilize deep learning to render outstanding results, and different object detection frameworks are discussed and analyzed with respect to their differences amongst each other. The comparisons are carried out based upon the criterion of accuracy and frames per second. The review starts with the explanation of CNNs and their advantages. Then we focus our attention on two types of object detection architectures and dive deep into their respective details. In the end, we draw a conclusion based on the facts detailed in the review.

Event-based cameras, also known as neuromorphic cameras, are bioinspired sensors able to perceive changes in the scene at high frequency with low power consumption. Becoming available only very recently, a limited amount of work addresses object detection on these devices. In this paper we propose two neural networks architectures for object detection: YOLE, which integrates the events into surfaces and uses a frame-based model to process them, and fcYOLE, an asynchronous event-based fully convolutional network which uses a novel and general formalization of the convolutional and max pooling layers to exploit the sparsity of camera events. We evaluate the algorithm with different extensions of publicly available datasets, and on a novel synthetic dataset.

Recently, self‑supervised representation learning relying on vast amounts of unlabeled data has been explored as a pre‑training method for autonomous driving. However, directly applying popular contrastive or generative methods to this problem is insufficient and may even lead to negative transfer. In this paper, we present AD‑L‑JEPA, a novel self‑supervised pre‑training framework with a joint embedding predictive architecture (JEPA) for automotive LiDAR object detection. Unlike existing methods, AD‑L‑JEPA is neither generative nor contrastive. Instead of explicitly generating masked regions, our method predicts Bird’s‑Eye‑View embeddings to capture the diverse nature of driving scenes. Furthermore, our approach eliminates the need to manually form contrastive pairs by employing explicit variance regularization to avoid representation collapse. Experimental results demonstrate consistent improvements on the LiDAR 3D object detection downstream task across the KITTI3D, Waymo, and ONCE datasets, while reducing GPU hours by 1.9×–2.7× and GPU memory by 2.8×–4× compared with the state-of-the-art method Occupancy-MAE. Notably, on the largest ONCE dataset, pre‑training on 100K frames yields a 1.61 mAP gain, better than in case of all the other methods pre‑trained on either 100K or 500K frames, and pre‑training on 500K frames yields a 2.98 mAP gain, better than in case of all the other methods pre‑trained on either 500K or 1M frames. AD‑L‑JEPA constitutes the first JEPA‑based pre‑training method for autonomous driving. It offers better quality, faster, and more GPU‑memory‑efficient self‑supervised representation learning.