Lightweight Object Detection Using Quantized YOLOv4-Tiny for Emergency Response in Aerial Imagery (2506.09299v1)

Abstract: This paper presents a lightweight and energy-efficient object detection solution for aerial imagery captured during emergency response situations. We focus on deploying the YOLOv4-Tiny model, a compact convolutional neural network, optimized through post-training quantization to INT8 precision. The model is trained on a custom-curated aerial emergency dataset, consisting of 10,820 annotated images covering critical emergency scenarios. Unlike prior works that rely on publicly available datasets, we created this dataset ourselves due to the lack of publicly available drone-view emergency imagery, making the dataset itself a key contribution of this work. The quantized model is evaluated against YOLOv5-small across multiple metrics, including mean Average Precision (mAP), F1 score, inference time, and model size. Experimental results demonstrate that the quantized YOLOv4-Tiny achieves comparable detection performance while reducing the model size from 22.5 MB to 6.4 MB and improving inference speed by 44\%. With a 71\% reduction in model size and a 44\% increase in inference speed, the quantized YOLOv4-Tiny model proves highly suitable for real-time emergency detection on low-power edge devices.

• The paper introduces a lightweight object detection framework using quantized YOLOv4-Tiny that reduces model size by 71% and boosts real-time performance for aerial emergency response.
• It leverages a custom dataset of over 10,000 annotated emergency images to optimize feature extraction and ensure efficient deployment on low-power edge devices.
• Quantization to INT8 precision significantly cuts power consumption while maintaining inference speed, making the approach ideal for resource-constrained, real-world emergency scenarios.

Lightweight Object Detection with Quantized YOLOv4-Tiny in Aerial Emergency Imagery

The paper "Lightweight Object Detection Using Quantized YOLOv4-Tiny for Emergency Response in Aerial Imagery" presents a thorough exploration of object detection techniques tailored for real-time applications, particularly within emergency response scenarios using aerial footage. The authors have chosen to focus on YOLOv4-Tiny, a compact version within the YOLO family, due to its suitability for deployment on low-power edge devices, such as Raspberry Pi systems, highlighting its balance of performance and efficiency.

The paper is grounded on a custom-created aerial imagery dataset specifically curated for emergency applications, showcasing 10,820 images annotated across various emergency instances. The dataset encompasses diverse elements like police vehicles, fire engines, and ambulances, providing critical visual cues for emergency detection tasks. The authors undertook the creation of this dataset given the dearth of publicly available sets specific to drone-view emergency imagery, marking a significant contribution to emergency management and response workflows.

Technical Highlights:

1. Model Architecture: YOLOv4-Tiny operates as a streamlined version of YOLOv4, utilizing CSPDarknet53-tiny as its backbone to ensure efficient feature extraction with reduced computational overhead. This model is optimized further by employing post-training quantization to compress weights to INT8 precision, effectively reducing its size to 6.4 MB from the original 22.5 MB. This method significantly enhances inference speed by 44%, a critical consideration for real-time applications.
2. Comparative Analysis: The quantized YOLOv4-Tiny was benchmarked against YOLOv5-small, another model in the lightweight category. Despite YOLOv5-small demonstrating slightly superior metrics such as mAP and F1 Score, YOLOv4-Tiny excelled in computational efficiency, offering faster inference times and reduced power consumption, attributes vital for deployment in constrained environments like aerial drones or mobile edge devices.
3. Quantization Impact: The transition from FP32 to INT8 precision through ONNX Runtime APIs resulted in substantial power efficiency gains, cutting average power usage from 33.8 W to 13.85 W. The slight increase in inference time post-quantization was offset by dramatic reductions in power consumption, establishing the model's suitability for environments where energy conservation is paramount.

Practical and Theoretical Implications:

The efficient implementation of quantized YOLOv4-Tiny demonstrates the feasibility of deploying powerful detection systems in resource-constrained settings. This paper extends the ongoing dialogue around model compression strategies, emphasizing the trade-offs between accuracy and energy efficiency—an area crucial to embedded systems development and real-world deployment in autonomous aerial surveillance tasks. Such advancements pave way for integrating AI-driven detection into broader applications beyond emergency response, potentially influencing sectors such as smart city infrastructure, environmental monitoring, and autonomous vehicular systems.

Future Directions:

The findings suggest promising avenues where further research can build on the quantization and deployment strategies outlined in the paper. Potential explorations may include further compression techniques like pruning, addressing class imbalance in datasets through innovative augmentation strategies, and the integration of transfer learning for cross-domain model applications. Additionally, enhancement of interpretability through explainable AI methods could augment model outputs, improving trust and transparency in high-stakes decision-making environments.

Overall, the paper profoundly impacts the paper of lightweight object detection in aerial imagery, showcasing the critical balance between AI sophistication and hardware constraints, opening pathways towards more efficient, robust, and energy-conscious real-time sensing applications.