Forget Less, Count Better: A Domain-Incremental Self-Distillation Learning Benchmark for Lifelong Crowd Counting | Newswise – Newswise

Forget Less, Count Better: A Domain-Incremental Self-Distillation Learning Benchmark for Lifelong Crowd Counting | Newswise – Newswise

 

Advancing Sustainable Urban Development Through Innovative Crowd Counting Technology

Introduction: Aligning Technological Innovation with Global Goals

A recent study from Fudan University, published in Frontiers of Information Technology & Electronic Engineering, introduces a novel framework for lifelong crowd counting. This technological advancement, titled the “Forget Less, Count Better” (FLCB) framework, directly addresses critical challenges in urban management and public safety, aligning with several United Nations Sustainable Development Goals (SDGs). By enhancing the ability of systems to accurately count crowds across diverse environments, this research contributes significantly to:

  • SDG 9: Industry, Innovation, and Infrastructure: Fostering innovation and building resilient infrastructure capable of managing complex urban systems.
  • SDG 11: Sustainable Cities and Communities: Making cities and human settlements inclusive, safe, resilient, and sustainable by providing tools for effective public space management and safety monitoring.

The Challenge: Overcoming Technological Barriers to Sustainable Urban Management

Catastrophic Forgetting in Crowd Counting Models

Traditional crowd counting models face a significant limitation known as “catastrophic forgetting.” When these models are trained incrementally on data from new and different domains (e.g., a different city square or type of event), they tend to lose the knowledge acquired from previous domains. This technical hurdle poses a substantial barrier to developing robust and scalable solutions for urban monitoring, directly impacting the ability to achieve SDG 11 targets for safe and efficient public spaces. A model that cannot generalize or adapt without forgetting past information is inherently unsustainable for real-world, dynamic urban environments.

The FLCB Framework: A Sustainable Solution for Lifelong Learning

Core Methodology and Contribution to SDGs

The proposed FLCB framework utilizes a domain-incremental self-distillation learning approach to create a model that continuously improves its performance across all domains. This method is a key innovation for sustainable infrastructure (SDG 9) for the following reasons:

  1. Knowledge Reusability: The framework uses knowledge distillation, allowing a model trained on new data to be guided by the knowledge from old domains without needing to store the old data. This reduces storage overhead and computational costs, making the technology more efficient and sustainable.
  2. Enhanced Adaptability: A balanced domain forgetting loss function ensures the model effectively learns new information while retaining old knowledge. This adaptability is crucial for creating resilient systems (SDG 9.1) that can be deployed in any urban domain for reliable person counting.
  3. Improved Public Safety and Planning (SDG 11): The final model’s ability to perform accurately across seen and unseen datasets directly supports the creation of safer public transport systems, better urban planning, and more effective crowd management during large-scale events, contributing to targets under SDG 11.2 (safe, affordable, accessible transport) and SDG 11.7 (safe, inclusive public spaces).

Performance Analysis and Generalization Capabilities

Empirical Results Supporting Sustainable Deployment

Comparative experiments demonstrate that the FLCB framework significantly mitigates the catastrophic forgetting phenomenon. The model achieved the lowest error rates (mMAE, mRMSE) and forgetting measures (nBwT) compared to classic lifelong learning methods. Furthermore, the FLCB framework exhibited strong generalization ability in both real-to-real and synthetic-to-real experiments, achieving lower prediction errors on entirely new, unseen datasets. Visualization of the density maps confirms that FLCB provides more accurate crowd density estimates, a critical function for applications in public health monitoring (SDG 3) and disaster response (SDG 11).

Future Directions for Enhanced Contribution to the SDGs

Addressing Limitations and Expanding Applications

The study acknowledges current limitations, such as difficulties with missing annotations and background noise, which must be addressed to maximize the technology’s utility in complex, real-world urban settings. Future research directions aimed at strengthening the framework’s contribution to the SDGs include:

  • Exploring efficient data sampling and self-supervised learning to reduce reliance on perfectly annotated data.
  • Improving robustness to environmental noise to ensure reliability in diverse urban conditions.
  • Extending the lifelong learning principles to other critical regression tasks relevant to sustainable development, such as traffic flow prediction or environmental monitoring.

By pursuing these advancements, the FLCB framework and similar innovations can become even more powerful tools in the global effort to build smarter, safer, and more sustainable cities for all.

SDGs Addressed or Connected to the Issues Highlighted in the Article

  1. SDG 9: Industry, Innovation, and Infrastructure

    • The article is centered on a technological innovation, a “domain-incremental self-distillation learning benchmark (FLCB) framework,” developed by researchers from Fudan University. This directly relates to fostering innovation and enhancing scientific research. The paper itself, published in Frontiers of Information Technology & Electronic Engineering, is a product of scientific research aimed at upgrading technological capabilities in the field of crowd counting.
  2. SDG 11: Sustainable Cities and Communities

    • The core application of the technology, “lifelong crowd counting,” is fundamental to urban management. Accurate person counting can be used to manage public spaces, optimize public transport, and ensure safety in crowded urban environments. The article mentions the goal is to have the “final model can be deployed to any domain for person counting,” which is a key tool for creating safer and more efficient cities.

Specific Targets Under Those SDGs Identified Based on the Article’s Content

  1. Target 9.5: Enhance scientific research, upgrade the technological capabilities of industrial sectors in all countries…encouraging innovation…

    • The entire article describes a research paper that introduces a new technological framework (FLCB). It explicitly aims to solve existing problems in crowd counting, such as “catastrophic forgetting and generalization ability issues,” thereby enhancing scientific research and upgrading the technology used for this task. The paper’s publication and its findings represent a direct contribution to this target.
  2. Target 11.7: By 2030, provide universal access to safe, inclusive and accessible, green and public spaces…

    • The technology for “lifelong crowd counting” can be directly applied to monitor and manage public spaces. The article highlights that the FLCB framework “estimates crowd density more accurately on both seen and unseen datasets.” This improved accuracy is crucial for authorities to manage crowd levels in parks, squares, and public venues to ensure they remain safe and do not become dangerously overcrowded.

Indicators Mentioned or Implied in the Article

  1. Technical Performance Metrics

    • The article explicitly mentions several quantitative indicators used to measure the performance and progress of the FLCB framework. These are direct measures of the technology’s effectiveness.
    • mMAE (mean Absolute Error) and mRMSE (mean Root Mean Squared Error): The article states that FLCB achieved the “lowest mMAE, mRMSE” values. These indicators measure the prediction error of the crowd counting model. A lower value signifies higher accuracy in counting, which is a direct measure of progress towards a more reliable tool for urban management (Target 11.7) and a more advanced technology (Target 9.5).
    • nBwT (normalized Backward Transfer): This is mentioned as another metric where FLCB scored the lowest. In the context of lifelong learning, this indicator measures how well the model retains knowledge from old tasks after learning new ones, directly addressing the “catastrophic forgetting phenomenon.” It serves as an indicator of the model’s stability and efficiency.
  2. Qualitative Performance Assessment

    • Visualization of prediction density maps: The article notes that “Visualization of prediction density maps shows that FLCB estimates crowd density more accurately.” This serves as a qualitative indicator of the technology’s improved performance and its ability to provide a more precise understanding of crowd distribution, which is essential for safety and planning.
  3. Innovation and Research Output

    • Publication of the research paper: The existence of the paper itself, “Forget less, count better: a domain-incremental self-distillation learning benchmark for lifelong crowd counting,” is an indicator of progress in scientific research as per Target 9.5.

SDGs, Targets, and Indicators Analysis

SDGs Targets Indicators
SDG 9: Industry, Innovation, and Infrastructure Target 9.5: Enhance scientific research, upgrade the technological capabilities of industrial sectors in all countries…encouraging innovation…
  • Publication of the research paper on the FLCB framework.
  • Demonstrated reduction in “catastrophic forgetting” via the nBwT metric.
  • Improved model performance shown by lower mMAE and mRMSE values compared to classic methods.
SDG 11: Sustainable Cities and Communities Target 11.7: By 2030, provide universal access to safe, inclusive and accessible, green and public spaces…
  • Lower prediction errors (mMAE, mRMSE) in crowd counting, indicating more reliable data for managing public space safety.
  • More accurate estimation of crowd density, confirmed by “Visualization of prediction density maps,” allowing for better crowd management.
  • The model’s “strong generalization ability,” allowing it to be deployed in various unseen domains (different public spaces) for person counting.

Source: newswise.com