Design and optimization of an embedded machine learning image processing system with a Linux SOC for large-scale livestock tallying

This thesis examines the challenges and solutions related to deploying machine learning models for livestock tracking and subsequent tallying on edge devices in remote areas. The resource demands of detection models typically necessitate cloud-based inference, which is impractical for remote livestock agricultural settings with limited infrastructure and resources. Various detection architectures and tracking algorithms are investigated for their capacity to operate in real-world constrained-resource environments to address the aforementioned drawback. Modern hardware and software elements that promote edge acceleration of machine learning algorithms are presented, building on the latest technologies to inform a prototype system. As limited research is available for the applicability of current detection and tracking solutions in the context of livestock tracking, a modular system design is taken to support continual experimentation and performance analysis of tracking components. An object detection dataset of cattle imagery is constructed and utilised as a benchmark dataset during the evaluation of the latest You Only Look Once (YOLO detection models, encompassing the YOLOv9, YOLOv10, and YOLOv11 and their size variants. NVIDIA’s TensorRT acceleration framework is used to optimise these custom-trained architectures for FP32, FP16, and INT8 quantization formats, enabling further performance assessment on the selected NVIDIA Jetson Orin Nano platform. The YOLOv9’s large model variant with FP16 quantization proved superior and achieved a mean average precision (mAP) across intersection over union thresholds between 0.50 to 0.95 (mAP50-95) of 0.788 at an inference speed of 19.4ms, followed by its medium variant with an accuracy of 0.784 at 17.9ms. YOLOv11’s x-variant obtained the highest mAP50 accuracy of 0.941 on the CattleTracker dataset, executing at 20.1ms. The ByteTrack, BotSort, and BoostTrack tracking algorithms were benchmarked on a separate cattle Multiple Object Tracking (MOT) dataset with YOLOv9l-FP16 as the detection component, where BotSort displayed the optimal trade-off between accuracy and speed. It achieved a Higher Order Tracking Accuracy (HOTA) of 0.82 at a processing time of 9.9ms, measured for the tracking component alone. ByteTrack obtained a throughput of 4ms, but at the cost of a lower 0.794 HOTA accuracy. BoostTrack produced the lowest HOTA metric of 0.782, with a significant number of identity switches on the custom CattleTracker tracking dataset. Using the top contenders of both experimentation phases, the YOLOv9l- FP16 and BotSort detection-and-tracker pair were tested in the field to generate a livestock tallying solution and verify the system’s capacity for real-world deployment.

Grantor dc:publisher

Stellenbosch : Stellenbosch University

Year dc:date.issued

2026

Author dc:creator

• Le Roux, Morne

Advisors dc:contributor.advisor

• Duckitt, W. D.
• Niesler, Thomas
• Wolhuter, Riaan

Language dc:language.iso

en

Repository record dc:identifier.uri

https://scholar.sun.ac.za/handle/10019.1/136291

OAI identifier oai:identifier

oai:scholar.sun.ac.za:10019.1/136291