Abstract
Aiming at the lightweight and real-time detection requirements of infrared pedestrians on edge devices, this paper proposes an improved infrared pedestrian recognition model based on YOLOv5n. The model effectively reduces the number of parameters by introducing the GhostConv and SlimNeck modules in the lightweight design. To enhance real-time performance, the model reduces the number of parameters and improves operation speed through the PAGCP pruning strategy. After experimental analysis, a balance between operation speed and accuracy is achieved. The results show that compared to the original model, the proposed model decreases the number of parameters by 1.52M, increases running speed by 21%, and decreases mAP.50 by only 0.4%. The running speed on Jetson Nano reaches 0.12s per image. The proposed model effectively ensures high detection accuracy under lightweight and real-time requirements, providing technical support for the deployment of edge devices in scenarios such as autonomous driving.
Introduction
Autonomous driving systems rely heavily on robust visual perception to ensure safety. However, standard RGB cameras often fail in challenging environmental conditions such as low light, heavy fog, or rain. Infrared (thermal) imaging provides a viable solution by detecting heat signatures, allowing for consistent performance regardless of ambient lighting.
Despite the advantages of infrared imaging, deploying high-performance deep learning models on edge devices—the low-power computers typically found in vehicles, such as the NVIDIA Jetson Nano—remains a significant challenge. These models are often computationally expensive, leading to high latency that is unacceptable for real-time safety applications. This research addresses the need for a lightweight, high-speed infrared pedestrian recognition model that maintains high accuracy while operating within the constraints of edge hardware.
Methodology
The proposed model is an optimized version of YOLOv5n, the most compact variant of the YOLOv5 family. The optimization focuses on three key areas: backbone efficiency, feature fusion, and structural pruning.
Model Architecture
The overall architecture is divided into three main components: the Backbone for feature extraction, the Slim Neck for feature fusion, and the Head for final detection.
GhostConv Optimization
Standard convolutional layers generate many redundant feature maps, which increases the computational load. To mitigate this, the authors integrated GhostConv into the backbone. GhostConv operates by using a small number of standard convolutions to produce "intrinsic" feature maps, which are then augmented with "ghost" features generated via computationally inexpensive linear transformations.
SlimNeck and GSConv
The "neck" of the network is responsible for fusing features across different scales. The researchers replaced the standard neck with a SlimNeck architecture utilizing GSConv and VoV-GSCSP modules. This design preserves the necessary cross-channel interactions required for accurate detection while significantly lowering the parameter count compared to traditional feature pyramid networks.
PAGCP Pruning
To further refine the model, the authors applied Performance-Aware Global Channel Pruning (PAGCP). Unlike naive pruning methods that might remove filters based solely on local importance, PAGCP evaluates the joint influence of convolutional kernels across the entire network. This global perspective allows for more aggressive "trimming" of redundant channels without incurring a substantial loss in detection accuracy.
Experimental Setup and Datasets
The model was trained and validated using the LLVIP (Low-light Visible-Infrared Paired) Dataset. The study focused exclusively on the infrared modality, utilizing 15,488 images captured across 26 synchronized scenes. The dataset was partitioned into an 8:2 ratio for training and validation purposes.
Results and Discussion
The optimized model was benchmarked against the standard YOLOv5n baseline and other modern architectures like YOLOv8.
- Efficiency Gains: The model's parameter count was reduced from 1.87M to just 0.35M, representing an 81% reduction in size.
- Real-Time Performance: On a standard CPU, the running speed increased by 21%. When deployed on the target NVIDIA Jetson Nano edge device, the model achieved an inference time of 0.12s per image (approximately 8 FPS).
- Accuracy Retention: Despite the massive reduction in complexity, the [email protected] only decreased by 0.4%, moving from 98.4% to 98.0%.
- Comparative Advantage: When compared to YOLOv8s, the proposed model uses only 3% of the parameters and is 72% faster on CPU while maintaining competitive detection quality.
Conclusion
This research demonstrates that through strategic architectural modifications—specifically GhostConv and SlimNeck—and intelligent pruning, it is possible to create highly efficient models for infrared pedestrian detection. While the current 8 FPS on a Jetson Nano is suitable for low-speed robotics or surveillance, future work will focus on multi-modal fusion (combining visible and infrared data) and further inference optimizations to reach the frame rates required for high-speed autonomous driving.
Summary updated