Student or educator? Sign up with your school email and get 6 months of Pro free.

The DOI matches its Crossref record; 58 of the 64 references it cites were found in a registry or an index; 3 of the 4 sources its claims cite were checked for support.

Edge YOLO: Real-Time Intelligent Object Detection System Based on Edge-Cloud Cooperation in Autonomous Vehicles

Open access
Liang, Siyuan, Wu, Hao, Zhen, Li, Hua, Qiaozhi, Garg, Sahil, Kaddoum, Georges, Hassan, Mohammad Mehedi, Yu, Keping·2022·IEEE Transactions on Intelligent Transportation Systems

Abstract

Driven by the ever-increasing requirements of autonomous vehicles, such as traffic monitoring and driving assistant, deep learning-based object detection (DL-OD) has been increasingly attractive in intelligent transportation systems. However, it is difficult for the existing DL-OD schemes to realize the responsible, cost-saving, and energy-efficient autonomous vehicle systems due to low their inherent defects of low timeliness and high energy consumption. In this paper, we propose an object detection (OD) system based on edge-cloud cooperation and reconstructive convolutional neural networks, which is called Edge YOLO. This system can effectively avoid the excessive dependence on computing power and uneven distribution of cloud computing resources. Specifically, it is a lightweight OD framework realized by combining pruning feature extraction network and compression feature fusion network to enhance the efficiency of multi-scale prediction to the largest extent. In addition, we developed an autonomous driving platform equipped with NVIDIA Jetson for system-level verification. We experimentally demonstrate the reliability and efficiency of Edge YOLO on COCO2017 and KITTI data sets, respectively. According to COCO2017 standard datasets with a speed of 26.6 frames per second (FPS), the results show that the number of parameters in the entire network is only 25.67 MB, while the accuracy (mAP) is up to 47.3%.

Overview

CoChat Summary

## TL;DR

Edge YOLO is a lightweight edge-cloud object detection system for autonomous vehicles that keeps real-time inference on Jetson edge devices while using the cloud for training and weight updates. The paper reports 47.3% mAP on COCO2017 at 26.6 FPS, with 25.27 MB model size and 8 million parameters.

## Key findings

  • On COCO2017, Edge YOLO reached 47.3% mAP at 26.6 FPS with a model size of 25.27 MB; the paper also states the network has 8 million parameters and about 10.2% of the original YOLOv4 structure.
  • On the KITTI 2D object dataset, Edge YOLO achieved 0.8212 mAP with 40.6 FPS on Jetson Xavier and 9.7 FPS on Jetson Nano, according to the comparison table.
  • Compared with YOLOV3, YOLOV4, MobileNetV3 SSD, and YOLOV3-Tiny, Edge YOLO is presented as a balance between speed and accuracy; on COCO2017 it is reported at 26.6 FPS versus 11.8, 5.2, 11.4, and 62.7 FPS for those baselines, respectively, with the corresponding mAP values listed in Table V.
  • The paper states average power consumption is 16.8 W and claims the edge-cloud cooperation scheme improves real-time performance while reducing dependence on cloud resources.

## Methods

  • The system is an edge-cloud cooperation (E-CC) design: the edge device performs real-time inference, while the cloud handles data upload, labeling, retraining, and weight updates.
  • The model is built by modifying YOLOv4: the paper says it trims the Backbone using channel pruning, simplifies the Neck by retaining SPP + FPN and removing PANet parts, and uses Leaky_ReLU plus TensorRT for acceleration.
  • Experiments use COCO2017 and KITTI 2D Object Dataset, with training in Ubuntu 18.04 using PyTorch 1.6 for MobileNetV3 SSD and Darknet for YOLO-based networks; evaluation uses mAP, FPS, BFLOPS, size, and IoU = 0.5.

## Limitations

  • The paper itself notes that in super multi-target detection, Edge YOLO is slightly inferior to YOLOV4 because of the limited number of network layers and parameters.
  • The text makes evident that reported performance depends on specific hardware and inference settings, including Jetson Xavier/Nano, TensorRT FP16, and the stated network environment; this limits direct generalization beyond the tested setup.
  • The paper states that mAP drops to a certain extent during FP16 accelerated inference on the large COCO2017 dataset.

Model-written summary. Check it against the paper before relying on it.

Integrity signals

Registry

Crossref · 10.1109/tits.2022.3158253

Cites

58 of 58 references verified0 citation errors · 0 not checkable

Statistics

Nothing to recompute.

Figures

Fig. 1 Three different structures for cloud and edge computing layers and related applications based on neural networks: E-CC, Cloud Computing, edge Computing and Fog Computing.
Fig. 1 Three different structures for cloud and edge computing layers and related applications based on neural networks: E-CC, Cloud Computing, edge Computing and Fog Computing.From the PDF
Fig. 2 Two different working paths for object detections based on edge devices: Our E-CC solution (Purple path) and Cloud computing solution (Green path).
Fig. 2 Two different working paths for object detections based on edge devices: Our E-CC solution (Purple path) and Cloud computing solution (Green path).From the PDF
Fig. 3 Overview of the proposed Edge YOLO framework. It contains the improved CSPDarknet, SPPNet, FPN and YOLO heads.
Fig. 3 Overview of the proposed Edge YOLO framework. It contains the improved CSPDarknet, SPPNet, FPN and YOLO heads.From the PDF
Fig. 4 Detailed block diagram of CSPNet features fusion strategy. In the backbone network of Edge YOLO, the feature map represented by Part 2 is first transited and finally merged with the Feature map represented by Part 1 to effectively reduce the repeated gradient information.
Fig. 4 Detailed block diagram of CSPNet features fusion strategy. In the backbone network of Edge YOLO, the feature map represented by Part 2 is first transited and finally merged with the Feature map represented by Part 1 to effectively reduce the repeated gradient information.From the PDF
Fig. 5 Comparison of three activation functions: Mish, Leaky_ReLU and ReLU.
Fig. 5 Comparison of three activation functions: Mish, Leaky_ReLU and ReLU.From the PDF
Fig. 6 Bounding boxes with offset and grid location prediction. We predicted the position of the center coordinate of the mask relative to the filter application.
Fig. 6 Bounding boxes with offset and grid location prediction. We predicted the position of the center coordinate of the mask relative to the filter application.From the PDF
Image 7
From the PDF
26×26
26×26From the PDF
52×52
52×52From the PDF
Fig. 8 FPN applied in the one-stage architecture. Build a module by combining horizontal joins and top-down paths by addition.
Fig. 8 FPN applied in the one-stage architecture. Build a module by combining horizontal joins and top-down paths by addition.From the PDF
In TABLE I, the filter, stride and Billion Floating-Point Operations Per Second (BFLOPS) of each layer are shown
In TABLE I, the filter, stride and Billion Floating-Point Operations Per Second (BFLOPS) of each layer are shownFrom the PDF
Fig. 9 Training results for loss function by two Datasets.
Fig. 9 Training results for loss function by two Datasets.From the PDF
Fig. 10 Test examples for common objects in COCO dataset detected by Edge YOLO. On the left is the original image, and on the right is the detected image.
Fig. 10 Test examples for common objects in COCO dataset detected by Edge YOLO. On the left is the original image, and on the right is the detected image.From the PDF
Image 14
From the PDF
Fig. 12 The results of KITTI 2D detection test set based on Edge YOLO, YOLOV4 and YOLOV3-Tiny models. Edge YOLO shows a high generalization performance, which improves the detectability of small targets and alleviate the risk of missed or false detection.
Fig. 12 The results of KITTI 2D detection test set based on Edge YOLO, YOLOV4 and YOLOV3-Tiny models. Edge YOLO shows a high generalization performance, which improves the detectability of small targets and alleviate the risk of missed or false detection.From the PDF
Fig. 13 Test results from KITTI 2D detection test set detected by Edge YOLO and the same dataset trained by cloud computing. Where (a), (c) and (e) represent the detection effect of the original Edge YOLO network located in the cloud. Meanwhile, (b), (d) and (f) represent the detection effect of Edge YOLO at the Edge end after TensorRT accelerated reasoning.
Fig. 13 Test results from KITTI 2D detection test set detected by Edge YOLO and the same dataset trained by cloud computing. Where (a), (c) and (e) represent the detection effect of the original Edge YOLO network located in the cloud. Meanwhile, (b), (d) and (f) represent the detection effect of Edge YOLO at the Edge end after TensorRT accelerated reasoning.From the PDF
Fig. 14 Number of uploaded pictures and delays calculated by three different approaches: Cloud computing, Edge-Cloud computing (Jetson Nano) and Edge-Cloud computing (Jetson Xavier Nx).
Fig. 14 Number of uploaded pictures and delays calculated by three different approaches: Cloud computing, Edge-Cloud computing (Jetson Nano) and Edge-Cloud computing (Jetson Xavier Nx).From the PDF
Image 18
From the PDF
Image 19
From the PDF
Image 20
From the PDF
Image 21
From the PDF
Image 22
From the PDF
Image 23
From the PDF
Image 24
From the PDF
Image 25
From the PDF

Verification

Claims the paper makes

3 Valid · 4 Not valid · 1 Not checked · 51 Own results and interpretations

15 more claims

Create a free account to see all 59 claims and every source verdict.

See all claims — free account
  • It may state two separate claims as one. 29
  • Probably not a substantial claim. 14
  • The quote only partly states this claim. 4

Create a free account to see these claims.

Claims other papers make about this one

Each is checked against this paper’s own text: does it say what the citing paper says it does?

Citations

Cites

Show all 64 references

Cited by

See all 367

Metadata

Authors
Liang, S., Wu, H., Zhen, L., Hua, Q., Garg, S., Kaddoum, G., Hassan, M. M., & Yu, K.
Affiliations
  • Xi’an University of Posts and Telecommunications (CN)
  • Hubei University of Arts and Science (CN)
  • École de Technologie Supérieure (CA)
  • King Saud University (SA)
  • Waseda University (JP)
Venue
IEEE Transactions on Intelligent Transportation Systems
Published
2022-12
Volume
23
Issue
12
Pages
25345-25360
ISSN
  • 1524-9050 (ISSN-L)
  • 1558-0016
OpenAlex
W4220827132
Cited by
367
References
64
Citations by year
2026: 88 · 2025: 148 · 2024: 75 · 2023: 51 · 2022: 5
FWCI
22.83
Type
journal article
Language
en
Open access
Locations
  • IEEE Transactions on Intelligent Transportation Systems (publishedVersion)
  • arXiv (Cornell University) (submittedVersion)
  • Espace ÉTS (ETS) (submittedVersion)
  • arXiv (Cornell University) (cc-by)
Funders
King Saud University, Riyadh, Saudi Arabia, through the Researchers Supporting Project (RSP 2022/18)
Topics
  • Physical Sciences › Computer Science › Computer Vision and Pattern Recognition › Advanced Neural Network Applications
  • Physical Sciences › Computer Science › Computer Networks and Communications › IoT and Edge/Fog Computing
  • Physical Sciences › Engineering › Automotive Engineering › Autonomous Vehicle Technology and Safety
Keywords
object detection, autonomous vehicles, lightweight neural networks, network pruning, feature fusion, multi-scale prediction, real-time detection, energy efficiency, NVIDIA Jetson, traffic monitoring, driving assistance

About these checks

Registry, retraction, Cites and statistics results are computed, not judged, with one exception: a printed reference that no index lists is read once by a model before it counts as a citation error. Claims are extracted by a model and kept only when their quote appears verbatim in the text; each cited source is then checked, with a model’s help, for whether it says what the claim relies on. Those verdicts can be wrong — read the claim’s quote and the cited source before relying on one.

The claim verdicts this report counts are experimental: their accuracy is not yet measured.

A language model drew the claims from the paper on 2026-09-28, under version 3 of the claim definition (model deepseek-v4.1-flash, prompt extract-v3).

A second model labelled each claim on 2026-09-28 (model jev-1.13-20260917, definition revision gold-guideline-v0.2).

Cited sources are checked with a language model’s help (gpt-6-sol).

Statistics and arithmetic are recomputed, not judged.

Confidence bands are shown only for fields calibrated on our annotated test set. None is calibrated yet, so no band is shown.

What counts as a claim ›How we measure confidence ›How we check this ›

Are you an author, or do you see a problem with this page? Request a review

How to cite

Cite CoChat for the checks on this page, and the paper for its own findings.

This page
CoChat. (2026). Citation check: Edge YOLO: Real-Time Intelligent Object Detection System Based on Edge-Cloud Cooperation in Autonomous Vehicles. https://app.cochat.ai/papers/10.1109/tits.2022.3158253
The paper
Liang, S., Wu, H., Zhen, L., Hua, Q., Garg, S., Kaddoum, G., Hassan, M. M., & Yu, K. (2022). Edge YOLO: Real-Time Intelligent Object Detection System Based on Edge-Cloud Cooperation in Autonomous Vehicles. IEEE Transactions on Intelligent Transportation Systems. https://doi.org/10.1109/tits.2022.3158253

Data sources

Bibliographic data: Semantic Scholar, OpenAlex, Crossref.

Contains information from the Semantic Scholar Academic Graph Datasets, made available under the ODC Attribution License (ODC-BY 1.0).

Semantic Scholar’s data is described in: Kinney, R., et al. (2023). The Semantic Scholar Open Data Platform. arXiv:2301.10140.