LSP-YOLO: Lightweight Posture Recognition

• The paper introduces a novel single-stage architecture that unifies keypoint estimation and posture classification for efficient real-time performance.
• It leverages a compact Light-C3k2 backbone with partial convolution and parameter-free SimAM attention to substantially reduce computational cost.
• The model achieves state-of-the-art accuracy on desktop setups while maintaining practical deployability on low-power embedded devices.

LSP-YOLO is a lightweight, single-stage convolutional neural network architecture designed specifically for efficient sitting posture recognition on embedded edge devices. Developed as an end-to-end solution that unifies keypoint estimation and posture classification, LSP-YOLO targets real-time applications with severe constraints on computational resources, such as smart classrooms, rehabilitation platforms, and human–computer interfaces. Its core innovations include the introduction of a compact Light-C3k2 backbone featuring partial convolution and parameter-free SimAM attention, as well as a direct, pointwise keypoint-to-class mapping in the recognition head. LSP-YOLO achieves state-of-the-art classification accuracy, extremely high throughput on desktop hardware, and practical deployability on low-power processors (Li et al., 18 Nov 2025).

1. Model Architecture and Design Principles

LSP-YOLO builds on the backbone–neck–head paradigm established in YOLOv11-Pose, but with explicit fusion of pose estimation and posture classification in a single forward pass. The model structure is as follows:

• Backbone: A stack of convolutional and Light-C3k2 modules replaces the conventional C3k2 blocks. The Spatial Pyramid Pooling Fast (SPPF) module enhances the receptive field for robust context capture.
• Neck: A PANet-style multi-scale fusion merges shallow spatial with deep semantic features across three scales, enabling the network to capture multi-level information crucial for keypoint localization and pose inference.
• Recognition Head (LSP-Head): For each output grid cell, the head jointly predicts confidence, regresses 11 upper-body keypoints, and classifies posture via a 1×1 convolution mapping the keypoint vector to six class logits, followed by a softmax.

Module Pipeline (Simplified)

Input → [Conv+Light‐C3k2]×n → SPPF → [Light‐C3k2 + up/down-sampling fusion] (neck) → LSP-Head → {confidence, keypoints, class-scores}

This single-stage approach eliminates the need for separate pose estimation pipelines, reducing both memory footprint and inference latency (Li et al., 18 Nov 2025).

2. Light-C3k2 Block: Partial Convolution and SimAM

The core backbone module, Light-C3k2, merges efficiency-centric and attention mechanisms:

• Partial Convolution (PConv): Rather than convolving all $c$ channels, PConv applies standard convolution only to a fraction $r$ (set to 0.5), passing the remaining channels via identity. The computational savings for a $k\times k$ kernel are

$\text{Relative FLOP reduction} = 1 - r^2$

yielding 75% reduction per 3×3 conv when $r=0.5$.

• Similarity-Aware Activation Module (SimAM): A parameter-free attention mechanism, SimAM computes an importance energy for each neuron $t$ by optimizing

$$E(w_t, b_t) = \frac{1}{M-1} \sum_{i\neq t} [-1 - (w_t x_i + b_t)]^2 + [1 - (w_t t + b_t)]^2 + \lambda w_t^2$$

with attention score

$a_t = \mathrm{sigmoid}(1/e_t^*)$

acting channel-wise and location-wise.

• Block Composition: Two Bottlenecks (k2), each starting with a PConv, followed by 1×1 convs for fusion, SimAM after each Bottleneck, and a residual connection encompassing both.

Light-C3k2 controls feature dimension via width multiplier $\alpha\in\{0.25, 0.5, 0.75, 1.0\}$, maintaining representation power while substantially reducing GFLOPs and memory requirements (Li et al., 18 Nov 2025).

3. Recognition Head and Losses

The LSP-Head handles all prediction targets per output grid cell:

• Keypoints-to-Class Mapping: The estimated keypoint vector $\hat K \in \mathbb{R}^D$ is transformed by a 1×1 conv to obtain six posture scores $s_i$, with softmax normalization:

$$S = \mathrm{Conv}_{1\times1}(\hat K), \quad \hat p_i = \frac{e^{s_i}}{\sum_{j=1}^6 e^{s_j}}$$

• Intermediate Supervision: Keypoint accuracy is enforced with an Object Keypoint Similarity (OKS) loss prior to class mapping, ensuring features support both regression and classification.

Loss Terms

Term	Loss Function	Purpose
Confidence	$$L_{\mathrm{conf}} = \sum_{n=1}^{N}\mathrm{BCE}(p_{\mathrm{conf}}^n,t_{\mathrm{conf}}^n)$$	Object presence
Keypoint	$$L_{\mathrm{oks}} = 1 - \frac{\sum_i \exp(-d_i^2/(2 s^2 k_i^2))\,\delta(v_i>0)}{\sum_i \delta(v_i>0)}$$	Keypoint accuracy
Classification	$$L_{\mathrm{cls}} = \sum_{n=1}^N \mathrm{BCE}(p_{\mathrm{cls}}^n,t_{\mathrm{cls}}^n)$$	Posture class accuracy

The aggregated loss is

$$L_{\mathrm{sum}} = \alpha L_{\mathrm{conf}} + \beta L_{\mathrm{oks}} + \gamma L_{\mathrm{cls}}$$

with $\alpha=1$, $\beta=12$, $\gamma=4$ (Li et al., 18 Nov 2025).

4. Dataset Construction and Augmentation

LSP-YOLO was trained and validated on a dedicated posture dataset with the following properties:

• Images: 5,000, annotated for six upper-body posture classes: Correct, LeanLeft, LeanRight, ChinSupport, OnDesk, HeadDown.
• Annotations: Each sample labeled with a bounding box (upper body), class, and 11 keypoints.
• Partitioning: 70% training, 15% validation, 15% testing.
• Augmentations: Random scaling, horizontal shift, HSV jitter, and random horizontal flips to bolster generalization (Li et al., 18 Nov 2025).

5. Training Process and Inference Results

• Training: Conducted for 300 epochs with batch size 32, learning rate annealed from 0.01 to $1\mathrm{e}{-4}$, image size 640×640, on AMD EPYC 7742 with dual RTX 3090 GPUs.
• Model Variants: The smallest, LSP-YOLO-n ($\alpha=0.25$, depth=0.33), contains 1.9 M parameters and requires 4.2 GFLOPs per inference.
• Performance on PC: Achieves 251 fps and 94.2% precision with a model size of 1.9 MB.
• Embedded Deployment: On the SV830C + GC030A platform (0.5 TOPS, 64 MB RAM, 640×480@30 fps camera), the 8-bit quantized model yields:
  • Preprocessing latency: 115 ms
  • Inference latency: 255 ms (≈4 fps)
  • Memory footprint: 22 MB
  • Accuracy: 91.7%
  • Model size: 2.2 MB

LSP-YOLO thus demonstrates both real-time throughput on desktop and practical, memory-constrained inference on edge accelerators (Li et al., 18 Nov 2025).

6. Computational Efficiency and Deployment Considerations

The model’s efficiency arises from design innovations:

• GFLOPs Reduction: The combination of PConv (reducing the dominant convolutional cost to 25%) and SimAM (no extra parameters) leads to a total GFLOP reduction of approximately 15–20% compared to the baseline, with greater than 96% retention of classification precision.
• Edge Compatibility: With model sizes near 2 MB, LSP-YOLO fits comfortably within the flash and RAM budgets of microcontroller- to mid-range systems.

Even the smallest variant consistently delivers 250 fps on high-end GPUs and approximately 4 fps on low-power hardware, validating its suitability for embedded applications (Li et al., 18 Nov 2025).

7. Applications, Limitations, and Directions for Further Research

Use Cases:

• Multi-student posture monitoring in smart classrooms
• Remote rehabilitation and posture correction feedback systems
• Human–computer interfaces leveraging posture-based control signals

Identified Limitations:

• Lower limb occlusion constrains reliable full-body posture estimation; expansion to 3D or multi-view sensing is a prospective solution.
• The current design processes single frames; temporal fusion with video streams could augment robustness and consistency.
• Scene-level multi-person counting is not addressed; future research can focus on dynamic keypoint grouping.
• Incorporation of self-supervised pretraining on large, unlabeled posture datasets could improve robustness to real-world variation.

LSP-YOLO thus provides a high-efficiency, deployable baseline for posture recognition research and applications, with multiple avenues for extension in both accuracy and scope (Li et al., 18 Nov 2025).