A Low-Complexity Human Head Detection Network Based on YOLOv6

Zhenyang Hu, Jiahao Yu, Gaofei Sun, Zhenjiang Qian, Xiaobing Xian, Miaomiao Zhu2023 · DOI

TL;DR

Researchers have developed a more efficient version of YOLOv6 specifically optimized for human head detection in crowded surveillance environments. By introducing a novel QIoU loss function and a lightweight decoupled head with attention mechanisms, the team reduced model parameters by 6.5% while simultaneously boosting detection accuracy, particularly for small and occluded targets.

What Problem Does This Paper Address?

Human head detection is a cornerstone of modern public safety and crowd management, used for everything from counting foot traffic to monitoring indoor security. However, it remains a "hard" problem for computer vision. Unlike full-body detection, heads are small, frequently occluded by other people, and their appearance changes drastically based on camera angle, lighting, and headwear (like hats or hairstyles).

While state-of-the-art models like YOLOv6 offer a great balance of speed and accuracy, they can still struggle in "dense" scenarios where heads overlap or appear as tiny clusters of pixels. The authors sought to create a "nano" scale model that is light enough for real-time edge deployment but robust enough to handle the complexities of a crowded classroom or a busy street.

Methods

The authors introduced two primary innovations to the YOLOv6-n (nano) baseline:

  1. QIoU (Quadruple Intersection over Union) Loss: Standard loss functions often struggle to help a model converge when targets are very small. The authors designed QIoU as an evolution of the DecIoU loss. It adds specific penalty terms for the X and Y center point coordinates of the bounding box. By forcing the model to minimize the distance between the predicted and true center points—alongside width and height—the model "snaps" onto targets faster and more accurately during training.
  2. Lightweight Decoupled Head: In YOLO models, the "head" is the part of the network that makes the final predictions. Inspired by the PPYOLOE architecture, the authors designed a lighter version that uses ESE (Effective Squeeze-and-Excitation) attention modules. This allows the network to focus more effectively on the spatial position of a head. By using 1x1 convolutions to extract features, they reduced the computational "weight" of the head while actually improving its ability to localize small objects.

Datasets

The model was trained and evaluated on the SCUT-HEAD dataset, a large-scale benchmark containing 4,405 images and over 111,000 annotated heads. The dataset is split into two parts:

  • Part A: Images from university classroom monitor videos (highly dense, consistent lighting).
  • Part B: Images crawled from the internet (diverse outdoor/indoor scenes, varying angles).

Key Insights and Ideas

  • Center Points are Crucial: The "aha moment" for the loss function was realizing that for small objects, even a slight shift in the center point leads to a massive drop in IoU. By explicitly penalizing center-point misalignment (the "Q" in QIoU), the model becomes much more stable in crowded scenes.
  • Attention over Raw Power: Instead of making the network deeper (which adds lag), the authors used ESE attention. This proves that where a model looks is often more important than how many layers it has, especially when the target (a head) has a very distinct but small footprint.
  • Efficiency Gains: The redesigned head proves that you can actually get "something for nothing." By simplifying the architecture with 1x1 kernels, they lowered the GFLOPs (computational cost) while increasing the Average Precision (AP).

Key Results

The improved YOLOv6-n model outperformed the standard baseline across the board:

  • Accuracy: Achieved 91.6% $AP_{50}$ and 52.2% AR (Average Recall), an improvement of 1.1% over the baseline using standard SIoU loss.
  • Small Target Performance: The model showed a significant jump in accuracy for small targets, reaching 47.6% $AR_s$ (a 1.5% increase over the baseline + QIoU).
  • Complexity: Reduced the number of parameters from 4.30M to 3.98M (a 6.5% reduction) and lowered computation from 11.06G to 10.43G GFLOPs.
  • Speed: Maintained a high inference speed of 103.3 FPS, making it ideal for real-time surveillance applications.

Limitations and Future Work

While the paper demonstrates strong results in dense environments, the authors note that occlusion and overlap remain the primary hurdles. When one head almost entirely covers another, even an optimized YOLOv6 can mistake the pair for a single object or background.

The paper does not explicitly detail performance in extreme low-light or "night vision" conditions, which would be a logical next step for a surveillance-focused model. Future work could involve further refining the attention mechanisms to better distinguish between overlapping boundaries in ultra-dense crowds.

Summary updated

Abstract

Indoor and outdoor surveillance have become essential aspects of public safety. In recent years, intelligent surveillance has made a lot of progress, with human head detection becoming a common technique in this field. However, there are some challenges in human head detection such as failure to detect targets, lighting conditions, and crowded areas. In this paper, we proposed an improved head network that decreases the number of YOLOv6-n model parameters while simultaneously increasing detection accuracy. Additionally, to accelerate model convergence, we designed a box regression loss function, QIoU, which converges in terms of the bounding box's width and height, as well as its X and Y axes center positions. Our improved algorithm achieves 91.6% AP and 52.2% AR on SCUT-HEAD dataset, respectively, which are higher than the baseline with SIoU loss function. Finally, our method also has better accuracy performance on small targets (i.e., 47.6 %).
BibTeX
@misc{hu2023low,
  author = {Zhenyang Hu and Jiahao Yu and Gaofei Sun and Zhenjiang Qian and Xiaobing Xian and Miaomiao Zhu},
  title = {{A Low-Complexity Human Head Detection Network Based on YOLOv6}},
  year = {2023},
  doi = {10.12792/icisip2023.036},
}
Figure