Optimization of Deep-Learning Detection of Humans in Marine Environment on Edge Devices (original) (raw)
Related papers
Marine Objects Detection Using Deep Learning on Embedded Edge Devices
2022
Artificial Intelligence techniques based on convolution neural networks (CNNs) are now dominant in the field of object detection and classification. The deployment of CNNs on embedded edge devices targeting real-time inference sets a challenge due to the limited computing resources and power budgets. Several optimization techniques such as pruning, quantization and use of light neural networks enable the realtime inference but at the cost of precision degradation. However, using efficient approaches to apply the optimization techniques at training and inference stages enable high inference speed with limited degradation of detection performance. In this paper, we revisit the problem of detecting and classifying maritime objects. We investigate different versions of the You Only Look Once (YOLO), a state-of-the-art deep neural network, for real-time object detection and compare their performance for the specific application of detecting maritime objects. The trained YOLO networks are efficiently optimized targeting three recent edge devices: Nvidia Jetson Xavier AGX, AMD-Xilinx Kria KV260 Vision AI Kit, and Movidius Myriad X VPU. The proposed deployments demonstrate promising results with an inference speed of 90 FPS and a limited degradation of 2.4% in mean average precision.
2020
In this paper Quantization effects are assessed for a real time Edge based person detection use case that is based on the use of a Raspberry Pi. TensorFlow architectures are presented that enable the use of real-time person detection on the Raspberry Pi. The model quantization is performed, performance of quantized models is analyzed, and worstcase performance is established for a number of deep learning object detection models that are capable of being deployed on the Pi for realtime applications. The study shows that the inference time for a suitably optimized TensorFlow enabled solution architecture is significantly lower than for an unquantized model with only slight cost implications in terms of accuracy when benchmarked against a desktop implementation. An industrial standard floor limit value of greater than 70% is achieved on the quantized models considered with a reduced detection time of less than 3ms. The Deep Neural Network model is trained using the INRIA Person Detecti...
Implementation of Deep-Learning-based Edge Computing for Maritime Vehicle Classification
2020
In recent years, Artificial Intelligence (AI) has revolutionized almost every field – from automotive industry to electrical industry, continuously affecting the dynamics of society and the lives of consumers. Among the countless applications of AI, AI at the edge provides the ideal solution for specific workload acceleration. Despite the development of various techniques for surveillance in maritime scenarios, automatic maritime vehicle classification of visual surveillance remains a challenge owing to the small dataset, as well as complex, unconstrained, and diverse nature of such scenarios. To date, only a few studies have investigated edge computing in maritime vehicle classification with edge device. In this paper, data augmentation is applied to deal with the small dataset problem because of a lack of images of military ships. Furthermore, the implementation of deep-learning-based edge computing in maritime vehicle classification through the use of NVIDIA Jetson Nano is proposed.
Real-Time Human Detection as an Edge Service Enabled by a Lightweight CNN
2018 IEEE International Conference on Edge Computing (EDGE)
Edge computing allows more computing tasks to take place on the decentralized nodes at the edge of networks. Today many delay sensitive, mission-critical applications can leverage these edge devices to reduce the time delay or even to enable real-time, online decision making thanks to their on-site presence. Human objects detection, behavior recognition and prediction in smart surveillance fall into that category, where a transition of a huge volume of video streaming data can take valuable time and place heavy pressure on communication networks. It is widely recognized that video processing and object detection are computing intensive and too expensive to be handled by resourcelimited edge devices. Inspired by the depthwise separable convolution and Single Shot Multi-Box Detector (SSD), a lightweight Convolutional Neural Network (L-CNN) is introduced in this paper. By narrowing down the classifier's searching space to focus on human objects in surveillance video frames, the proposed L-CNN algorithm is able to detect pedestrians with an affordable computation workload to an edge device. A prototype has been implemented on an edge node (Raspberry PI 3) using openCV libraries, and satisfactory performance is achieved using realworld surveillance video streams. The experimental study has validated the design of L-CNN and shown it is a promising approach to computing intensive applications at the edge.
Optimizing Real-Time Object Detection on Edge Devices
The issue of object detection in remote surveillance using edge devices presents a complex scenario, largely as a result of the limitations inherent in edge computing settings and the requirements for instantaneous data processing. Current video surveillance systems exhibit proficient video capture functionalities; however, data analysis at the server level is impeded by constraints in transmission power and the availability of cloud computing resources. Consequently, Internet of Things (IoT) devices are primarily relegated to the role of data acquisition. Our study proposes a novel fusion of ResNet18, K-Means clustering, and int8 quantization over tinyML, compressing the model to <100KB and enabling sub-1mW consumption on edge devices. This methodology extends the viability of deploying sophisticated machine-learning models on microcontrollers powered by coin cells, broadening their applicability in various settings for object detection. Employing int8 quantization, our model attains a notable improvement in latency by 45%, coupled with a 70% reduction in RAM consumption and a 65% decrease in flash storage. This research delves into the significance of optimization in the process of choosing latency-efficient deep neural network (DNN) models for various edge computing configurations, emphasizing the delicate balance between hardware capabilities and optimization approaches. Subsequent research efforts will concentrate on refining quantization algorithms to further mitigate the precision differential in computational models.
2024
This research conducts a comparative analysis of Faster R-CNN and YOLOv8 for real-time detection of fishing vessels and fish in maritime surveillance. The study underscores the significance of this investigation in advancing fisheries monitoring and object detection using deep learning. With a clear focus on comparing the performance of Faster R-CNN and YOLOv8, the research aims to elucidate their effectiveness in realtime detection, emphasizing the relevance of such capabilities in fisheries management. By conducting a thorough literature review, the study establishes the current state-ofthe-art in object detection, particularly within the context of fisheries monitoring, while discussing existing methods, challenges, and limitations. The findings of this study not only shed light on the superiority of YOLOv8 in precise detection but also highlight its potential impact on maritime surveillance and the protection of marine resources.
Deep learning has facilitated major advancements in various fields including image detection. This paper is an exploratory study on improving the performance of Convolutional Neural Network (CNN) models in environments with limited computing resources, such as the Raspberry Pi. A pretrained state-of-art algorithm for doing near-real time object detection in videos, YOLO ("You-Only-Look-Once") CNN model, was selected for evaluating strategies for optimizing the runtime performance. Various performance analysis tools provided by the Linux kernel were used to measure CPU time and memory footprint. Our results show that loop parallelization, static compilation of weights, and flattening of convolution layers reduce the total runtime by 85% and reduce memory footprint by 53% on a Raspberry Pi 3 device. These findings suggest that the methodological improvements proposed in this work can reduce the computational overload of running CNN models on devices with limited computing resources.
Proceedings of the 26th Irish Conference on Artificial Intelligence and Cognitive Science, 2018
This research paper investigates the running of object detection algorithms on low-end devices to detect individuals in images while leveraging cloud-based services to provide facial verification of the individuals detected. The performance of three computer vision object detection algorithms that utilize Convolutional Neural Networks (CNN) are compared: SSD MobileNet, Inception v2 and Tiny YOLO along with three cloud-based facial verification services: Kairos, AmazonWeb Service Rekognition (AWS) and Microsoft Azure Vision API. The results contribute to the limitations of running CNN based algorithms to solve vision tasks on low-end devices and highlights the limitations of using such devices and models in an application domain such as a home-security solution.
Efficient Ship Detection System for Maritime Surveillance Using Deep Learning Approach
International Journal For Multidisciplinary Research, 2024
The usage of drones in maritime surveillance is a very effective means to observe ships. As the ship detection is vital for applications such as port monitoring, cross-border surveillance, Moreover, it is not only vital for maritime surveillance but also environmental conservation. Despite its effectiveness, the drone-captured images present itself with its own challenges, specifically when leveraging on a custom dataset which is tailored for specific application. To address the issues, we propose a novel ship detection approach called YOLOv9 with Adan optimizer (YOLOv9-Adan) which is more accurate to increase the efficiency of maritime surveillance. The YOLOv9-Adan model integrates the robust object detection capabilities of the adaptive learning capabilities of the Adan (ADAptive Nesterov momentum algorithm) optimiser, trained on the drone image dataset comprising 3200 images of maritime scenes and ship types in drone view. Our model is trained on a drone-image dataset comprising 3200 images of maritime scenes and ship types in drone views collected from various sources. The experimental results show that our approach using the YOLOv9-Adan model achieves 65.5% mAP, which exceeds the mAP of YOLOv9 by 4.3%. Additionally, this article also provides a comparative analysis of our model YOLOv9-Adan with other existing models in literature with consistently surpassing existing approaches.
Applied Sciences
Human detection is a special application of object recognition and is considered one of the greatest challenges in computer vision. It is the starting point of a number of applications, including public safety and security surveillance around the world. Human detection technologies have advanced significantly in recent years due to the rapid development of deep learning techniques. Despite recent advances, we still need to adopt the best network-design practices that enable compact sizes, deep designs, and fast training times while maintaining high accuracies. In this article, we propose ReSTiNet, a novel compressed convolutional neural network that addresses the issues of size, detection speed, and accuracy. Following SqueezeNet, ReSTiNet adopts the fire modules by examining the number of fire modules and their placement within the model to reduce the number of parameters and thus the model size. The residual connections within the fire modules in ReSTiNet are interpolated and fine...