Detection of Lower Body for AGV Based on SSD Algorithm with ResNet

Xinbiao Gao, Junhua Xu, Chuan Luo, Jun Zhou, Panling Huang, Jianxin DengSensors 2022 · DOI

TL;DR

This paper introduces R-SSD, an improved object detection algorithm for identifying the lower body of humans, crucial for autonomous guided vehicles (AGVs). By replacing the VGG16 backbone with ResNet50, increasing input resolution, and employing multi-scale feature maps, R-SSD achieves a significant 7% mAP improvement over the original SSD, reaching 85.1% mAP.

What Problem Does This Paper Address?

Autonomous Guided Vehicles (AGVs) are increasingly used in industrial settings for logistics and transportation. To ensure efficient and safe operation, AGVs need to accurately track and relocate themselves within complex environments. A key challenge is their limited perceptual range, often focused on areas near the ground. Detecting the lower body of humans is vital for AGVs to navigate safely around people, preventing collisions and optimizing movement. Traditional object detection methods, while powerful, may struggle with the specific challenges of detecting human lower bodies in diverse industrial conditions, such as varying illumination, scale, and potential occlusions.

Methods

The core of this work is the development of an improved Single Shot MultiBox Detector (SSD) algorithm, termed R-SSD. The authors leverage the strengths of two established deep learning architectures:

  • SSD: A popular one-stage object detection framework known for its balance between speed and accuracy.
  • ResNet50: A powerful convolutional neural network architecture that utilizes residual connections to enable deeper networks and better feature extraction, mitigating the degradation problem often seen in very deep networks.

The R-SSD algorithm makes the following key modifications:

  • Backbone Replacement: The VGG16 backbone of the original SSD is replaced with ResNet50. This enhances the model's feature extraction capabilities, leading to improved detection accuracy.
  • Increased Input Resolution: The input image resolution is increased from the standard 300x300 to 448x448 pixels. This allows for better extraction of features from a larger image range, particularly beneficial for detecting objects at different distances.
  • Multi-Scale Feature Maps: The R-SSD utilizes six feature maps for detection, allowing it to capture objects at various scales. This is crucial for detecting human lower bodies, which can appear at different sizes depending on their distance from the camera.
  • Bounding Box Clustering: The authors employ a clustering approach based on YOLO9000 to define custom bounding box priors. This is done to better match the aspect ratios and sizes of human lower bodies, improving detection efficiency and accuracy.

Datasets

The study utilizes a custom dataset collected using a Daheng image industrial camera.

  • Dataset Size: 1132 images were used for training, with a total of 1580 labeled lower body instances. This dataset includes both complex and simple images, with a focus on factory workshop environments.
  • Annotation: The images were annotated using the open-source software "LabelImg," with the lower body labeled as a single category. The dataset specifically excludes data with obvious occlusions like skirts and includes images with partial occlusions to enhance robustness.

Key Insights and Ideas

  • ResNet's Superiority for Feature Extraction: Replacing VGG16 with ResNet50 significantly improves the feature extraction capabilities of the SSD model, leading to a noticeable increase in detection accuracy. This highlights the importance of choosing an appropriate backbone network for specific tasks.
  • Input Resolution Matters: Increasing the input resolution from 300x300 to 448x448 pixels demonstrably improves detection accuracy. This suggests that for tasks requiring fine-grained detection of objects like human lower bodies, higher resolution inputs can provide richer information.
  • Data Augmentation and Clustering are Crucial: The paper shows that both expanding the dataset with diverse scenarios (complex and simple backgrounds) and employing custom bounding box clustering contribute to improved detection performance. This underscores the importance of careful dataset preparation and tailored anchor box strategies.
  • Robustness in Diverse Conditions: The R-SSD model demonstrates good performance across various challenging conditions, including changes in illumination, object scale, and occlusion, indicating its practical applicability in real-world industrial environments.

Key Results

  • mAP Improvement: The R-SSD model achieved a mean Average Precision (mAP) of 85.1% on the test set.
  • Comparison to Original SSD: This represents a significant improvement of 7% mAP compared to the original SSD model (78.1% mAP).
  • Impact of ResNet: Using ResNet50 as the backbone improved the mAP from 78.1% to 80.7% (a 2.6% increase).
  • Impact of Input Size: Increasing input size to 448x448 improved mAP to 83.0%.
  • Impact of Data Expansion: Adding simple background images further boosted mAP to 84.5%.
  • Impact of Bounding Box Clustering: The final R-SSD with clustering achieved the highest mAP of 85.1%.
  • High Confidence: The detection confidence in practical applications reached over 99%, laying a strong foundation for subsequent tracking and relocation tasks.

Limitations and Future Work

  • Dataset Specificity: While the custom dataset is tailored for industrial AGV scenarios, the authors acknowledge that in future work, they plan to train their method on public datasets. This will help to further validate its generalizability and robustness across a wider range of environments.
  • Data Filling Strategy: The authors suggest that for different scenarios, a corresponding "data filling" strategy can be developed to achieve stable and reliable practical applications. This implies that further research into domain adaptation or specialized data augmentation techniques might be beneficial.

Summary updated

Abstract

Detection of human lower body provides an implementation idea for the automatic tracking and accurate relocation of automatic vehicles. Based on traditional SSD and ResNet, this paper proposes an improved detection algorithm R-SSD for human lower body detection, which utilizes ResNet50 instead of VGG16 to improve the feature extraction level of the model. According to the application of acquisition equipment, the model input resolution is increased to 448 × 448 and the model detection range is expanded. Six feature maps of the updated resolution network are selected for detection and the lower body image dataset is clustered into five categories for aspect ratio, which are evenly distributed to each feature detection map. The experimental results show that the model R-SSD detection accuracy after training reaches 85.1% mAP. Compared with the original SSD, the detection accuracy is improved by 7% mAP. The detection confidence in practical application reaches more than 99%, which lays the foundation for subsequent tracking and relocation for automatic vehicles.
BibTeX
@inproceedings{gao2022detection,
  author = {Xinbiao Gao and Junhua Xu and Chuan Luo and Jun Zhou and Panling Huang and Jianxin Deng},
  title = {{Detection of Lower Body for AGV Based on SSD Algorithm with ResNet}},
  booktitle = {Sensors},
  year = {2022},
  doi = {10.3390/s22052008},
}
Figure