Intelligent bear deterrence system based on computer vision: Reducing human-bear conflicts in remote areas

Jiawei Yi; John A. Kupfer; Pengyu Chen; Teng Fei; Yi Li; Yunyan Du

arxiv: 2503.23178 · v3 · submitted 2025-03-29 · 💻 cs.CV

Intelligent bear deterrence system based on computer vision: Reducing human-bear conflicts in remote areas

Pengyu Chen , Teng Fei , John A. Kupfer , Yunyan Du , Jiawei Yi , Yi Li This is my paper

Pith reviewed 2026-05-22 22:26 UTC · model grok-4.3

classification 💻 cs.CV

keywords bear deterrencecomputer visionedge AIhuman-wildlife conflictTibetan PlateauYOLOv5solar-powered IoT

0 comments

The pith

A solar-powered edge-AI system detects bears and activates spray within 0.2 seconds in remote off-grid areas.

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The paper develops and evaluates a complete bear-deterrence device that runs on solar power and local edge processing, with no reliance on cellular networks or mains electricity. It trains a YOLOv5-MobileNet model on a custom 1,243-image wildlife dataset and integrates the detector with a spray actuator. Laboratory trials report 97.2 percent activation accuracy across 100 simulated approaches, while a 30-day field deployment in Qinghai recorded three correct deterrences and zero false triggers. The authors conclude that the combination of low-power hardware and reliable detection offers a scalable, non-lethal method for reducing human-bear conflicts on the Tibetan Plateau.

Core claim

The integrated YOLOv5-MobileNet detector on a low-power edge board, paired with a solar-powered spray unit, achieves mean average precision of 91.4 percent and recall of 93.6 percent on the compiled wildlife dataset and delivers 97.2 percent accurate, sub-0.2-second spray activation in controlled tests plus three verified deterrence events in a 30-day real-world trial without false activations.

What carries the argument

YOLOv5-MobileNet model running on a low-power edge AI board that triggers a solar-powered bear-spray actuator when a bear is detected in the camera feed.

If this is right

Continuous off-grid operation becomes feasible for remote conflict zones.
Non-lethal, automatic deterrence reduces both human injury risk and the need for lethal control of bears.
The same hardware-software stack can be redeployed at additional sites without new infrastructure.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

The approach could be tested on other large mammals that cause similar conflicts in remote areas.
Adding low-cost environmental sensors might further reduce false triggers without raising power draw.
Wider replication would require confirming that the 1,243-image training set covers the full range of local lighting and animal postures.

Load-bearing premise

A 30-day trial that recorded only three bear events is enough to show the system will work reliably over longer periods and under changing weather, bear densities, and animal behaviors.

What would settle it

A multi-month deployment in the same region that records either repeated false activations or missed bear approaches under varied weather would show the performance metrics do not generalize.

Figures

Figures reproduced from arXiv: 2503.23178 by Jiawei Yi, John A. Kupfer, Pengyu Chen, Teng Fei, Yi Li, Yunyan Du.

**Figure 1.** Figure 1: Method Framework 3.1. Bear Detection Model In this phase, we collected wildlife images, annotated them with bounding boxes, and fine-tuned YOLOv5 on the MaixHub platform to obtain our bear detection model. 3.1.1. Data Collection We gathered over 1,000 wildlife images for training, including more than 600 bear images. To prepare for potential misclassifications, we also collected images of other animals com… view at source ↗

**Figure 2.** Figure 2: Dataset Distribution 3.1.2. Model Training We fine-tuned the YOLOv5 (You Only Look Once, version 5) model with MobileNet as its backbone to ensure efficient and accurate bear detection on low-power hardware [24, 25]. YOLOv5 is a well-established object detection framework known for balancing processing speed and detection accuracy, making it particularly suitable for real-time embedded applications. Mobi… view at source ↗

**Figure 3.** Figure 3: Bear Detection Workflow The key training parameters used in the fine-tuning process are summarized in [PITH_FULL_IMAGE:figures/full_fig_p008_3.png] view at source ↗

**Figure 4.** Figure 4: mAP value during training process Additionally, the model attained an F1 score of 94.7%, which balances precision and recall to provide a comprehensive assessment of the model’s performance. The false positive rate was low at 3.79%, indicating that only a small fraction of non-bear instances were misclassified as bears. This low rate is vital for practical deployment since it minimizes erroneous alerts and… view at source ↗

**Figure 5.** Figure 5: Video misclassification rates of common objects on the Tibetan Plateau [PITH_FULL_IMAGE:figures/full_fig_p011_5.png] view at source ↗

**Figure 6.** Figure 6: Location of Devices Solar panels power the entire system, ensuring continuous functionality in off-grid areas like Zadoi County [PITH_FULL_IMAGE:figures/full_fig_p012_6.png] view at source ↗

**Figure 7.** Figure 7: Internal components of the system Over a 30-day testing period, three successful Tibetan brown bear deterrence events were documented by a monitoring camera linked to the bear spray module. Remarkably, all events were recorded by the same unit, which had been strategically positioned near a water channel in a residential village—highlighting the camera’s effectiveness in a high-risk location. Of these, t… view at source ↗

**Figure 8.** Figure 8: A Bear Deterrence Event Captured by Monitoring Camera (bear highlighted [PITH_FULL_IMAGE:figures/full_fig_p014_8.png] view at source ↗

read the original abstract

Human-bear conflicts on the Tibetan Plateau threaten both local livelihoods and the conservation of Tibetan brown bears (Ursus arctos pruinosus). To address this challenge, we developed a low-power, network-independent deterrence system that combines computer vision with Internet of Things (IoT) hardware. The system integrates a YOLOv5-MobileNet detection model deployed on a low-power edge artificial intelligence (AI) board with a solar-powered bear spray device. We compiled a data set of 1,243 wildlife images (including 795 bears with 100 infrared captures for nighttime detection, plus other common objects and animals such as mastiffs, yaks, humans, and vehicles), from which 80% were used for training and 20% for validation. Validation showed robust performance (mean average precision = 91.4%, recall = 93.6%). In 100 controlled activation tests involving simulated approaches by bears, humans, and other animals, the spray deployed within 0.2 seconds of detection with 97.2% accuracy, confirming timely and reliable responses. A 30-day field trial in Zadoi County, Qinghai Province, China, recorded 3 successful deterrence events without false activations. By using energy-efficient components and ensuring continuous and stable system operation, this solution provides a practical, sustainable, and scalable approach to mitigating human-bear conflicts, effectively enhancing human safety and bear conservation in remote areas without network or grid coverage.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Desk Editor's Note private letter to a colleague

A working edge-AI bear-spray prototype with solid lab metrics but only three field events over 30 days, which undercuts the scalability claims.

read the letter

The paper describes a solar-powered, network-free system that runs YOLOv5-MobileNet on an edge board to detect bears and trigger spray. They assembled 1243 images (including infrared night shots and local animals like yaks and mastiffs), trained on 80% and validated on 20% to reach 91.4% mAP and 93.6% recall. In 100 controlled tests the actuator fired in 0.2 seconds at 97.2% accuracy. A 30-day deployment in Zadoi County logged three correct activations and zero false positives. That specific hardware-plus-actuator combination for Tibetan Plateau conditions is not in the cited prior work, so the application itself is new. The engineering choices for low power and off-grid operation are sensible and the controlled-test numbers line up with the claims. The dataset and split are clearly described, which helps reproducibility on that part. The main limitation is the field trial size. Three events give little information on performance across weather changes, different bear densities, or rare failure modes such as occlusion or spray misalignment. The abstract calls the approach practical, sustainable, and scalable, but the evidence stays narrow and short-term. No long-duration logs or detailed error analysis appear. This is useful reading for groups doing applied computer vision for wildlife management or remote conservation hardware. A reader who needs an example of edge deployment in harsh settings will find concrete details. It is coherent on its own terms and shows honest empirical work, so it should go to peer review rather than desk reject; referees will likely ask for more field data or a narrower scope statement.

Referee Report

1 major / 1 minor

Summary. The paper claims to have developed a low-power, network-independent bear deterrence system combining a YOLOv5-MobileNet detection model on an edge AI board with a solar-powered spray device. Using a custom dataset of 1,243 wildlife images (80/20 train/validation split), it reports mAP of 91.4% and recall of 93.6% on validation, 97.2% accuracy across 100 controlled activation tests, and three successful deterrences with zero false positives during a 30-day field trial in Zadoi County, concluding that the system provides a practical, sustainable, and scalable solution for mitigating human-bear conflicts in remote off-grid areas.

Significance. If the reported performance generalizes, the work demonstrates a viable engineering integration of edge computer vision with autonomous physical deterrence for wildlife management in challenging environments. The empirical results from both controlled tests and initial field deployment, along with emphasis on energy efficiency and independence from infrastructure, offer concrete data points useful for applied AI in conservation contexts.

major comments (1)

[Abstract (field trial description)] Abstract (field trial description): The assertion of a 'practical, sustainable, and scalable approach' depends on the 30-day field trial, which recorded only three successful deterrence events. This limited sample provides insufficient evidence to establish robustness to variable weather, bear densities, seasonal changes, or edge cases such as occlusions and low-light infrared failures, directly weakening support for the long-term reliability and scalability claims.

minor comments (1)

[Abstract (model description)] The abstract refers to 'YOLOv5-MobileNet' without specifying the exact architecture variant or training hyperparameters; adding these details in the methods would improve reproducibility.

Simulated Author's Rebuttal

1 responses · 0 unresolved

We thank the referee for their constructive feedback. We address the single major comment below.

read point-by-point responses

Referee: [Abstract (field trial description)] Abstract (field trial description): The assertion of a 'practical, sustainable, and scalable approach' depends on the 30-day field trial, which recorded only three successful deterrence events. This limited sample provides insufficient evidence to establish robustness to variable weather, bear densities, seasonal changes, or edge cases such as occlusions and low-light infrared failures, directly weakening support for the long-term reliability and scalability claims.

Authors: We agree that the 30-day field trial constitutes a limited sample and does not by itself demonstrate robustness across the full range of environmental variables, seasonal changes, or edge cases listed. The abstract's concluding claims draw on the combination of the controlled 100-test results (97.2 % accuracy), the 1,243-image validation metrics, and the low-power, solar, network-independent system architecture. Nevertheless, to address the concern directly, we will revise the abstract to qualify the language, replacing the final sentence with a more measured statement that the field results provide initial evidence of practical deployment. We will also add an explicit limitations paragraph in the discussion section that notes the small number of field events and outlines the need for longer-term trials. revision: yes

Circularity Check

0 steps flagged

No circularity: purely empirical measurements with no derivations or fitted predictions

full rationale

The paper reports an empirical pipeline: compilation of a 1243-image dataset, 80/20 train/validation split, YOLOv5-MobileNet training yielding mAP 91.4% and recall 93.6%, 100 controlled activation tests (97.2% accuracy), and a 30-day field deployment recording 3 events. No equations, parameter fits, uniqueness theorems, or self-citations are invoked as load-bearing steps. All performance numbers are direct outputs of the described experiments rather than predictions derived from the same data by construction. The central claim therefore rests on external validation data, not on any reduction to its own inputs.

Axiom & Free-Parameter Ledger

0 free parameters · 0 axioms · 0 invented entities

The work relies on standard assumptions of computer vision pipelines and hardware reliability without introducing new free parameters, axioms, or invented entities beyond the described system components.

pith-pipeline@v0.9.0 · 5812 in / 1066 out tokens · 37166 ms · 2026-05-22T22:26:52.814660+00:00 · methodology

discussion (0)

Reference graph

Works this paper leans on

25 extracted references · 25 canonical work pages · 1 internal anchor

[1]

Y. e. a. Dai, Mitigation strategies for human–tibetan brown bear (ur- sus arctos pruinosus) conflicts in the hinterland of the qinghai-tibetan plateau, Animals 12 (11) (2022). doi:10.3390/ani12111422

work page doi:10.3390/ani12111422 2022
[2]

Y. e. a. Dai, Conflicts of human with the tibetan brown bear (ursus arctos pruinosus) in the sanjiangyuan region, china, Global Ecology and Conservation 22 (2020) e01039. doi:10.1016/j.gecco.2020.e01039

work page doi:10.1016/j.gecco.2020.e01039 2020
[3]

Y. e. a. Dai, Human-carnivore conflicts and mitigation options in qinghai province, china, Journal for Nature Conservation 53 (2020) 125776.doi: 10.1016/j.jnc.2019.125776

work page doi:10.1016/j.jnc.2019.125776 2020
[4]

Y. Dai, D. Li, F. Liu, Y. Zhang, Summary comments on human-bear conflict mitigation measures and implications to sanjiangyuan national park, Acta Ecologica Sinica 39 (22) (2019) 8310–8318. doi:10.5846/ stxb201904060672

work page 2019
[5]

W. e. a. Tang, Meeting china’s electricity demand with renewable energy over tibetan plateau, Science Bulletin 68 (1) (2023) 39–42. doi:10. 1016/j.scib.2022.12.012

work page 2023
[6]

H. e. a. Xufang, The status and characteristics of, and solutions to, human-tibetan brown bear conflicts in the qinghai province, Acta The- riologica Sinica 38 (1) (2018) 28. doi:10.16829/j.slxb.150106

work page doi:10.16829/j.slxb.150106 2018
[7]

R. K. Choudhary, Internet of things: Wildlife conservation and its chal- lenges, Asian Journal of Computer Science and Technology 9 (1) (2020). doi:10.51983/ajcst-2020.9.1.2156

work page doi:10.51983/ajcst-2020.9.1.2156 2020
[8]

S. e. a. Sorensen, Deep learning for polar bear detection, in: Image Analysis, Springer International Publishing, 2017, pp. 457–467. doi: 10.1007/978-3-319-59126-1_38

work page doi:10.1007/978-3-319-59126-1_38 2017
[9]

A. e. a. Delplanque, Wildlife detection, counting and survey using satel- lite imagery: Are we there yet?, GIScience & Remote Sensing 61 (1) (2024) 2348863. doi:10.1080/15481603.2024.2348863. 18

work page doi:10.1080/15481603.2024.2348863 2024
[10]

A. M. e. a. Roy, Wildect-yolo: An efficient and robust computer vision-based accurate object localization model for automated endan- gered wildlife detection, Ecological Informatics 75 (2023) 101919. doi: 10.1016/j.ecoinf.2022.101919

work page doi:10.1016/j.ecoinf.2022.101919 2023
[11]

W. e. a. Yang, A forest wildlife detection algorithm based on improved yolov5s, Animals 13 (19) (2023). doi:10.3390/ani13193134

work page doi:10.3390/ani13193134 2023
[12]

S. e. a. Ward, Autonomous uavs wildlife detection using thermal imag- ing, predictive navigation and computer vision, in: 2016 IEEE Aerospace Conference, 2016, pp. 1–8. doi:10.1109/AERO.2016.7500671

work page doi:10.1109/aero.2016.7500671 2016
[13]

J. O. e. a. Chelotti, Livestock feeding behaviour: A review on automated systems for ruminant monitoring, Biosystems Engineering 246 (2024) 150–177. doi:10.1016/j.biosystemseng.2024.08.003

work page doi:10.1016/j.biosystemseng.2024.08.003 2024
[14]

M. e. a. Li, Social and cultural aspects of human–wildlife conflicts: Un- derstanding people’s attitudes to crop-raiding animals and other wildlife in agricultural systems of the tibetan plateau, Integrative Conservation 2 (4) (2023) 214–225. doi:10.1002/inc3.30

work page doi:10.1002/inc3.30 2023
[15]

H. Guo, Resolving human–bear conflict on the qinghai-tibetan plateau, Dialogue Earth https://dialogue.earth/en/nature/ resolving-human-bear-conflict-on-the-qinghai-tibetan-plateau/ (2021)

work page 2021
[16]

F. e. a. Liu, An integrated systems approach to plateau ecosystem management—a scientific application in qinghai and tibet plateau, In- formation Systems Frontiers 17 (2) (2015) 337–350. doi:10.1007/ s10796-012-9406-5

work page 2015
[17]

M. Foggin, Pastoralists and wildlife conservation in western china: Col- laborative management within protected areas on the tibetan plateau, Pastoralism: Research, Policy and Practice 2 (1) (2012) 17. doi: 10.1186/2041-7136-2-17

work page doi:10.1186/2041-7136-2-17 2012
[18]

T. Lu, L. Huntsinger, Managing human–wildlife conflict on the tibetan plateau, Ecosystem Health and Sustainability 9 (2023) 0023. doi:10. 34133/ehs.0023. 19

work page 2023
[19]

T. e. a. Zhang, Research advances in conservation and management of endangered mammals on the qinghai-tibet plateau, Acta Theriologica Sinica 42 (5) (2022) 490. doi:10.16829/j.slxb.150696

work page doi:10.16829/j.slxb.150696 2022
[20]

E. D. Ayele, N. Meratnia, P. J. M. Havinga, Towards a new opportunistic iot network architecture for wildlife monitoring system, 2018 9th IFIP International Conference on New Technologies, Mobility and Security (NTMS) (2018) 1–5 doi:10.1109/NTMS.2018.8328721

work page doi:10.1109/ntms.2018.8328721 2018
[21]

Thangavel, C

S. Thangavel, C. S. Shokkalingam, The iot based embedded system for the detection and discrimination of animals to avoid human–wildlife conflict, Journal of Ambient Intelligence and Humanized Computing 13 (6) (2022) 3065–3081. doi:10.1007/s12652-021-03141-9

work page doi:10.1007/s12652-021-03141-9 2022
[22]

A. R. e. a. Elias, Where’s the bear? automating wildlife image processing using iot and edge cloud systems, in: Proceedings of the Second Inter- national Conference on Internet-of-Things Design and Implementation, 2017, pp. 247–258. doi:10.1145/3054977.3054986

work page doi:10.1145/3054977.3054986 2017
[23]

com/ai/en/maixhub/index.html (2025)

Sipeed, Introduction to maixhub—sipeed wiki, https://wiki.sipeed. com/ai/en/maixhub/index.html (2025)

work page 2025
[24]

A. G. e. a. Howard, Mobilenets: Efficient convolutional neural networks for mobile vision applications, arXiv preprint arXiv:1704.04861 (2017)

work page internal anchor Pith review Pith/arXiv arXiv 2017
[25]

G. e. a. Jocher, ultralytics/yolov5: V7.0 - yolov5 sota realtime instance segmentation, https://doi.org/10.5281/zenodo.7347926 (2022). 20

work page doi:10.5281/zenodo.7347926 2022

[1] [1]

Y. e. a. Dai, Mitigation strategies for human–tibetan brown bear (ur- sus arctos pruinosus) conflicts in the hinterland of the qinghai-tibetan plateau, Animals 12 (11) (2022). doi:10.3390/ani12111422

work page doi:10.3390/ani12111422 2022

[2] [2]

Y. e. a. Dai, Conflicts of human with the tibetan brown bear (ursus arctos pruinosus) in the sanjiangyuan region, china, Global Ecology and Conservation 22 (2020) e01039. doi:10.1016/j.gecco.2020.e01039

work page doi:10.1016/j.gecco.2020.e01039 2020

[3] [3]

Y. e. a. Dai, Human-carnivore conflicts and mitigation options in qinghai province, china, Journal for Nature Conservation 53 (2020) 125776.doi: 10.1016/j.jnc.2019.125776

work page doi:10.1016/j.jnc.2019.125776 2020

[4] [4]

Y. Dai, D. Li, F. Liu, Y. Zhang, Summary comments on human-bear conflict mitigation measures and implications to sanjiangyuan national park, Acta Ecologica Sinica 39 (22) (2019) 8310–8318. doi:10.5846/ stxb201904060672

work page 2019

[5] [5]

W. e. a. Tang, Meeting china’s electricity demand with renewable energy over tibetan plateau, Science Bulletin 68 (1) (2023) 39–42. doi:10. 1016/j.scib.2022.12.012

work page 2023

[6] [6]

H. e. a. Xufang, The status and characteristics of, and solutions to, human-tibetan brown bear conflicts in the qinghai province, Acta The- riologica Sinica 38 (1) (2018) 28. doi:10.16829/j.slxb.150106

work page doi:10.16829/j.slxb.150106 2018

[7] [7]

R. K. Choudhary, Internet of things: Wildlife conservation and its chal- lenges, Asian Journal of Computer Science and Technology 9 (1) (2020). doi:10.51983/ajcst-2020.9.1.2156

work page doi:10.51983/ajcst-2020.9.1.2156 2020

[8] [8]

S. e. a. Sorensen, Deep learning for polar bear detection, in: Image Analysis, Springer International Publishing, 2017, pp. 457–467. doi: 10.1007/978-3-319-59126-1_38

work page doi:10.1007/978-3-319-59126-1_38 2017

[9] [9]

A. e. a. Delplanque, Wildlife detection, counting and survey using satel- lite imagery: Are we there yet?, GIScience & Remote Sensing 61 (1) (2024) 2348863. doi:10.1080/15481603.2024.2348863. 18

work page doi:10.1080/15481603.2024.2348863 2024

[10] [10]

A. M. e. a. Roy, Wildect-yolo: An efficient and robust computer vision-based accurate object localization model for automated endan- gered wildlife detection, Ecological Informatics 75 (2023) 101919. doi: 10.1016/j.ecoinf.2022.101919

work page doi:10.1016/j.ecoinf.2022.101919 2023

[11] [11]

W. e. a. Yang, A forest wildlife detection algorithm based on improved yolov5s, Animals 13 (19) (2023). doi:10.3390/ani13193134

work page doi:10.3390/ani13193134 2023

[12] [12]

S. e. a. Ward, Autonomous uavs wildlife detection using thermal imag- ing, predictive navigation and computer vision, in: 2016 IEEE Aerospace Conference, 2016, pp. 1–8. doi:10.1109/AERO.2016.7500671

work page doi:10.1109/aero.2016.7500671 2016

[13] [13]

J. O. e. a. Chelotti, Livestock feeding behaviour: A review on automated systems for ruminant monitoring, Biosystems Engineering 246 (2024) 150–177. doi:10.1016/j.biosystemseng.2024.08.003

work page doi:10.1016/j.biosystemseng.2024.08.003 2024

[14] [14]

M. e. a. Li, Social and cultural aspects of human–wildlife conflicts: Un- derstanding people’s attitudes to crop-raiding animals and other wildlife in agricultural systems of the tibetan plateau, Integrative Conservation 2 (4) (2023) 214–225. doi:10.1002/inc3.30

work page doi:10.1002/inc3.30 2023

[15] [15]

H. Guo, Resolving human–bear conflict on the qinghai-tibetan plateau, Dialogue Earth https://dialogue.earth/en/nature/ resolving-human-bear-conflict-on-the-qinghai-tibetan-plateau/ (2021)

work page 2021

[16] [16]

F. e. a. Liu, An integrated systems approach to plateau ecosystem management—a scientific application in qinghai and tibet plateau, In- formation Systems Frontiers 17 (2) (2015) 337–350. doi:10.1007/ s10796-012-9406-5

work page 2015

[17] [17]

M. Foggin, Pastoralists and wildlife conservation in western china: Col- laborative management within protected areas on the tibetan plateau, Pastoralism: Research, Policy and Practice 2 (1) (2012) 17. doi: 10.1186/2041-7136-2-17

work page doi:10.1186/2041-7136-2-17 2012

[18] [18]

T. Lu, L. Huntsinger, Managing human–wildlife conflict on the tibetan plateau, Ecosystem Health and Sustainability 9 (2023) 0023. doi:10. 34133/ehs.0023. 19

work page 2023

[19] [19]

T. e. a. Zhang, Research advances in conservation and management of endangered mammals on the qinghai-tibet plateau, Acta Theriologica Sinica 42 (5) (2022) 490. doi:10.16829/j.slxb.150696

work page doi:10.16829/j.slxb.150696 2022

[20] [20]

E. D. Ayele, N. Meratnia, P. J. M. Havinga, Towards a new opportunistic iot network architecture for wildlife monitoring system, 2018 9th IFIP International Conference on New Technologies, Mobility and Security (NTMS) (2018) 1–5 doi:10.1109/NTMS.2018.8328721

work page doi:10.1109/ntms.2018.8328721 2018

[21] [21]

Thangavel, C

S. Thangavel, C. S. Shokkalingam, The iot based embedded system for the detection and discrimination of animals to avoid human–wildlife conflict, Journal of Ambient Intelligence and Humanized Computing 13 (6) (2022) 3065–3081. doi:10.1007/s12652-021-03141-9

work page doi:10.1007/s12652-021-03141-9 2022

[22] [22]

A. R. e. a. Elias, Where’s the bear? automating wildlife image processing using iot and edge cloud systems, in: Proceedings of the Second Inter- national Conference on Internet-of-Things Design and Implementation, 2017, pp. 247–258. doi:10.1145/3054977.3054986

work page doi:10.1145/3054977.3054986 2017

[23] [23]

com/ai/en/maixhub/index.html (2025)

Sipeed, Introduction to maixhub—sipeed wiki, https://wiki.sipeed. com/ai/en/maixhub/index.html (2025)

work page 2025

[24] [24]

A. G. e. a. Howard, Mobilenets: Efficient convolutional neural networks for mobile vision applications, arXiv preprint arXiv:1704.04861 (2017)

work page internal anchor Pith review Pith/arXiv arXiv 2017

[25] [25]

G. e. a. Jocher, ultralytics/yolov5: V7.0 - yolov5 sota realtime instance segmentation, https://doi.org/10.5281/zenodo.7347926 (2022). 20

work page doi:10.5281/zenodo.7347926 2022