CoChat

Edge-Computing Video Analytics for Real-Time Traffic Monitoring in a Smart City

Open access

The DOI matches its Crossref record; 42 of the 42 references it cites were found in a registry or an index; 4 of the 12 sources its claims cite were checked for support.

Abstract

The increasing development of urban centers brings serious challenges for traffic management. In this paper, we introduce a smart visual sensor, developed for a pilot project taking place in the Australian city of Liverpool (NSW). The project’s aim was to design and evaluate an edge-computing device using computer vision and deep neural networks to track in real-time multi-modal transportation while ensuring citizens’ privacy. The performance of the sensor was evaluated on a town center dataset. We also introduce the interoperable Agnosticity framework designed to collect, store and access data from multiple sensors, with results from two real-world experiments.

Overview

CoChat Summary

## TL;DR

The paper presents an edge-computing visual sensor for real-time traffic monitoring from CCTV feeds that processes video on-device and transmits only metadata to preserve privacy. It reports validation on the Oxford Town Center dataset and two real-world deployments in indoor and outdoor settings, including a one-week Liverpool deployment with 20 sensors.

## Key findings

  • On the Oxford Town Center Dataset, the sensor was evaluated on the first 4,500 frames (3 minutes) containing 230 annotated pedestrians, and achieved a mean accuracy of 69% with a median relative error of 33%.
  • The sensor tended to underestimate detections; the paper states it was more prone to false negatives than false positives, and that accuracy dropped in larger crowds because of occlusions.
  • In a 10-minute real-world deployment, the system showed stable disk, memory, and temperature usage, while GPU use occasionally dropped from 100% to under 40% because the CPU-based tracking step became a bottleneck.
  • In a 1-hour indoor experiment, the sensor detected and tracked 631 people, and during a fire alarm event it detected no one, which the authors interpret as evidence that the building was effectively evacuated.
  • In the Liverpool outdoor deployment, the paper reports counts over one week (20 Feb 2019 to 27 Feb 2019) and notes 20,399 unique objects were detected on 23 February 2019; the results showed daily activity peaks around noon for pedestrians and around 09:00 and 16:00 for vehicles.

## Methods

  • Design: an edge-computing visual sensor built on an NVIDIA Jetson TX2 and a Pycom LoPy 4 for LoRaWAN, running detection and tracking onboard and transmitting only metadata over Ethernet or LoRaWAN.
  • Data/sample: validation used the Oxford Town Center Dataset; deployment tests included a 10-minute CCTV-connected system check, a 1-hour indoor monitoring experiment, and a one-week outdoor Liverpool deployment with 20 sensors.
  • Main analysis: YOLO V3 was used for object detection and SORT for tracking; performance was assessed with accuracy, relative error, and FPS, alongside system and network utilization measurements.

## Limitations

  • The paper explicitly states that the current implementation of SORT on the CPU was a bottleneck and suggests moving it to the GPU in future work.
  • The authors say that a more thorough investigation of the error rate would benefit the algorithm’s accuracy, indicating the evaluation is not fully exhaustive.
  • Evident from the text: the validation dataset was limited to the first 4,500 frames of one sequence and included only 230 annotated pedestrians, so the accuracy result is based on a relatively narrow test set.

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

Integrity signals

Registry

Crossref · 10.3390/s19092048

Cites

42 of 42 references verified0 citation errors · 0 not checkable

Statistics

Nothing to recompute.

Figures

Figure 1 A map of the town center. (left) Locations of the 20 visual sensors. It is expected that the live data produced from those location will help urban planners to update the city’s mobility plan. (right) Co-location of two CCTV cameras and an air quality (in purple) on a pole.
Figure 1 A map of the town center. (left) Locations of the 20 visual sensors. It is expected that the live data produced from those location will help urban planners to update the city’s mobility plan. (right) Co-location of two CCTV cameras and an air quality (in purple) on a pole.CC BY · PubMed Central
Figure 2 The smart visual sensor, outside and inside.
Figure 2 The smart visual sensor, outside and inside.CC BY · PubMed Central
Figure 3 Simplified diagram of the sensor. The core of the sensor is the Jetson TX2 handling the video analytics, the USB and Ethernet connectivity and powering the LoPy 4 and the OLED screen. The Jetson TX2 is in turn powered by the PSU. The LoPy is used for LoRaWAN-based transmission and the screen to display basic information about the sensor’s status.
Figure 3 Simplified diagram of the sensor. The core of the sensor is the Jetson TX2 handling the video analytics, the USB and Ethernet connectivity and powering the LoPy 4 and the OLED screen. The Jetson TX2 is in turn powered by the PSU. The LoPy is used for LoRaWAN-based transmission and the screen to display basic information about the sensor’s status.CC BY · PubMed Central
Figure 4 Activity flow chart of the sensor.
Figure 4 Activity flow chart of the sensor.CC BY · PubMed Central
Figure 5 The 3U server version of the visual sensor hosting 15 NVIDIA Jetson TX2 units, each of these computing modules are able to process in real time one live video feed from a CCTV camera.
Figure 5 The 3U server version of the visual sensor hosting 15 NVIDIA Jetson TX2 units, each of these computing modules are able to process in real time one live video feed from a CCTV camera.CC BY · PubMed Central
Figure 6 Architecture of YOLO V3 for object detection. A picture is passed through a fully convolutional neural network of 406 hidden layers in order to predict box coordinates and class probabilities at three different scales (large, medium, small). A non-maximum suppression algorithm is then applied to only retain the category and coordinates with the higher score.
Figure 6 Architecture of YOLO V3 for object detection. A picture is passed through a fully convolutional neural network of 406 hidden layers in order to predict box coordinates and class probabilities at three different scales (large, medium, small). A non-maximum suppression algorithm is then applied to only retain the category and coordinates with the higher score.CC BY · PubMed Central
Figure 7 The general architecture of the project. The Agnosticity software stack relies on well established open-source software. The data collection and access is ensured by the open-source implementation of the OM2M standard.
Figure 7 The general architecture of the project. The Agnosticity software stack relies on well established open-source software. The data collection and access is ensured by the open-source implementation of the OM2M standard.CC BY · PubMed Central
Figure 8 The web-based interactive dashboard used to represent the collected data from the different sensors deployed for the Liverpool project. The interface is responsive and can be used both on desktop and mobile browsers.
Figure 8 The web-based interactive dashboard used to represent the collected data from the different sensors deployed for the Liverpool project. The interface is responsive and can be used both on desktop and mobile browsers.CC BY · PubMed Central
Figure 9 Kernel density estimation of the accuracy (left) and percentage deviation (right) computed across the 4500 frames of the Oxford dataset.
Figure 9 Kernel density estimation of the accuracy (left) and percentage deviation (right) computed across the 4500 frames of the Oxford dataset.CC BY · PubMed Central
Figure 10 Evolution of the number of frame per seconds processed (FPS) by sensor (blue) and the number of detected objects (red) over time. It can be seen that the FPS are higher when the number of detection is lower. The drop in FPS is mainly due to the SORT algorithm as there are more objects to be tracked, a task not taking advantage of the CUDA cores available on the Jetson TX2.
Figure 10 Evolution of the number of frame per seconds processed (FPS) by sensor (blue) and the number of detected objects (red) over time. It can be seen that the FPS are higher when the number of detection is lower. The drop in FPS is mainly due to the SORT algorithm as there are more objects to be tracked, a task not taking advantage of the CUDA cores available on the Jetson TX2.CC BY · PubMed Central
Figure 11 Evolution of the accuracy (blue lines, ideal is 1.0) in accordance with the ground truth (orange lines) over time. Accuracy is better with small groups and decreases with large crowds.
Figure 11 Evolution of the accuracy (blue lines, ideal is 1.0) in accordance with the ground truth (orange lines) over time. Accuracy is better with small groups and decreases with large crowds.CC BY · PubMed Central
Figure 12 Number of ground truth detection against the number of detection. We observe that the relationship is linear, meaning that the algorithm manages to capture trends.
Figure 12 Number of ground truth detection against the number of detection. We observe that the relationship is linear, meaning that the algorithm manages to capture trends.CC BY · PubMed Central
Figure 13 A 15 min monitoring of the CPU, GPU, memory and disk usages, average temperature (top) and network utilization (bottom) during a real world deployment of the sensor. Over this period, 280 unique objects have been detected.
Figure 13 A 15 min monitoring of the CPU, GPU, memory and disk usages, average temperature (top) and network utilization (bottom) during a real world deployment of the sensor. Over this period, 280 unique objects have been detected.CC BY · PubMed Central
Figure 14 Plot of the number of people detected inside a building over one hour. The two peaks correspond to the start and end of the fire alarm event.
Figure 14 Plot of the number of people detected inside a building over one hour. The two peaks correspond to the start and end of the fire alarm event.CC BY · PubMed Central
Figure 15 Trajectories followed by the individuals detected and tracked by the sensor. Each of the 631 lines represent one individual.
Figure 15 Trajectories followed by the individuals detected and tracked by the sensor. Each of the 631 lines represent one individual.CC BY · PubMed Central
Figure 16 Heat map of the maximum number of individual detected in the field of view of the sensor.
Figure 16 Heat map of the maximum number of individual detected in the field of view of the sensor.CC BY · PubMed Central
Figure 17 The sensor is located at the center of the city, next to a pedestrian street (highlighted in red). The field of view of the camera is represented in blue.
Figure 17 The sensor is located at the center of the city, next to a pedestrian street (highlighted in red). The field of view of the camera is represented in blue.CC BY · PubMed Central
Figure 18 Number of pedestrian, vehicle and bicycle detected by the sensor between 20 February 2019 and 27 February 2019. Each data point represents the number of objects of a specific type detected over the last minute.
Figure 18 Number of pedestrian, vehicle and bicycle detected by the sensor between 20 February 2019 and 27 February 2019. Each data point represents the number of objects of a specific type detected over the last minute.CC BY · PubMed Central
Figure 19 Number of detection pedestrians (green), bicycles (red), and vehicles (yellow) detected hourly on 23 February 2019.
Figure 19 Number of detection pedestrians (green), bicycles (red), and vehicles (yellow) detected hourly on 23 February 2019.CC BY · PubMed Central
Figure 20 Plot of the pixel coordinates (X,Y) of the detected pedestrians (blue) or bicycles (orange) in the frame for the day 23 February 2019. Each dot represents the centroid of the bounding box corresponding to an item that has been detected at those coordinates.
Figure 20 Plot of the pixel coordinates (X,Y) of the detected pedestrians (blue) or bicycles (orange) in the frame for the day 23 February 2019. Each dot represents the centroid of the bounding box corresponding to an item that has been detected at those coordinates.CC BY · PubMed Central
Figure 21 Trajectories of pedestrians and bicycles within the frame during the day of 23 February 2019. Two flows of pedestrians are visible.
Figure 21 Trajectories of pedestrians and bicycles within the frame during the day of 23 February 2019. Two flows of pedestrians are visible.CC BY · PubMed Central

Verification

Claims the paper makes

4 Valid · 2 Not valid · 4 Not checked · 69 Own results and interpretations

30 more claims

Create a free account to see all 79 claims and to verify the 1 reference still open.

See all claims — free account
  • It may state two separate claims as one. 22
  • Probably not a substantial claim. 30
  • The quote only partly states this claim. 1

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 42 references

Research works printed but not linked: 1

  1. Khadir, K., Monteil, T., and Medjiah, S. (2018). IDE-OM2M: A framework for IoT applications using the the development of OM2M platform. Proceedings on the International Conference on Internet Computing, The Steering Committee of The World Congress in Computer Science, Computer Engineering & Applied Computing.

Cited by

Items 101–200 of 205

Metadata

Authors
Barthélemy, J., Verstaevel, N., Forehead, H., & Perez, P.
Affiliations
University of Wollongong (AU)
Corresponding author
Barthélemy, Johan
Venue
Sensors
Publisher
MDPI AG
Published
2019-05-02
Volume
19
Issue
9
Pages
2048
ISSN
1424-8220 (ISSN-L)
OpenAlex
W2942806183
Cited by
203
References
41
Citations by year
2026: 13 · 2025: 31 · 2024: 26 · 2023: 35 · 2022: 35 · 2021: 33 · 2020: 22 · 2019: 8
FWCI
10.95
Type
journal article
Language
en
Open access
Locations
  • Sensors (publishedVersion, cc-by)
  • PubMed Central (publishedVersion, cc-by)
  • PubMed (publishedVersion)
  • Research Online (University of Wollongong) (submittedVersion)
  • HAL (submittedVersion)
  • DOAJ (DOAJ: Directory of Open Access Journals) (submittedVersion, cc-by-sa)
  • Figshare (submittedVersion)
  • Repository of the University of Namur (submittedVersion, other-oa)
Funders
  • Nvidia
  • Australian Government
Topics
  • Physical Sciences › Computer Science › Computer Vision and Pattern Recognition › Video Surveillance and Tracking Methods
  • Physical Sciences › Computer Science › Computer Networks and Communications › IoT and Edge/Fog Computing
  • Physical Sciences › Computer Science › Computer Vision and Pattern Recognition › Advanced Neural Network Applications
Keywords
edge computing, video analysis, real-time traffic monitoring, smart cities, computer vision, deep neural network, citizen privacy, sensor interoperability

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.

A language model drew the claims from the paper on 2026-10-10, under version 3 of the claim definition (model deepseek-v4.1-flash, prompt extract-v3).

A second model labelled each claim on 2026-10-10 (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

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.