💻 computer science

Knowledge Evolution and AI Paradigm Shifts in Pedestrian Detection: A Bibliometric and Main Path Analysis

This study integrates bibliometric analysis with an AI paradigm perspective to map the structural evolution of pedestrian detection research, revealing that the field's transition from handcrafted features to deep learning and transformers occurs through gradual knowledge integration rather than abrupt shifts, thereby offering a novel theoretical framework for understanding AI development.

Original authors: Vo Thanh Kiet, Rene Jaros, Boris Pustejovsky, Jakub Stefansky, Ondrej Svoboda, Minh Ly Duc, Petr Bilik, Radek Martinek

Published 2026-06-29

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Vo Thanh Kiet, Rene Jaros, Boris Pustejovsky, Jakub Stefansky, Ondrej Svoboda, Minh Ly Duc, Petr Bilik, Radek Martinek

Original paper licensed under CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine the field of pedestrian detection (teaching computers to spot people in video) as a massive, bustling library that has been growing for over 30 years. This paper doesn't just read the books; it acts like a super-librarian who maps out the entire building, traces how ideas have traveled from one shelf to another, and identifies which books are the "bestsellers" that everyone keeps referencing.

Here is a simple breakdown of what the authors, Vo Thanh Kiet and his team, discovered:

1. The Library's Growth Spurt

For a long time (1990–2004), this library was quiet. Only a few people were writing about how to spot people using "hand-crafted" rules (like drawing specific shapes to find a person).

The Shift: Around 2010, the library exploded. It's like a sudden rush of new students. By 2020, the number of new books published every year jumped from a few dozen to over 500.
The Cause: This boom happened because the library switched from "hand-drawing rules" to Deep Learning (letting the computer learn the rules by looking at millions of photos).

2. Who is Writing the Books? (The Collaboration Map)

The authors mapped out who is talking to whom.

The Big Players: China is the giant in this room, writing more than half of all the books. The United States is the second largest, followed by India, Japan, and Germany.
The Social Network: While China writes the most, the UK and France are the most social. They write fewer books, but they collaborate with many different countries.
The Problem: The library is a bit fragmented. Many groups of researchers are like cliques in high school—they talk to each other a lot but rarely talk to the groups in other countries or schools. The authors suggest we need more "mixers" to connect these isolated groups.

3. The "Main Path" of Knowledge

Imagine knowledge as a river flowing from the past to the future. The authors used a special tool to trace the Main Path—the strongest, most important current of ideas.

The Old River (1990s–2000s): The water flowed through "Handcrafted Features." Scientists manually designed rules (like "people have two legs") to find pedestrians.
The Transition (The Middle Ground): The river didn't dry up and start a new one immediately. There was a swampy middle zone where old rules and new AI methods coexisted. Scientists were testing both at the same time, bridging the gap.
The New River (2015–Present): The water now flows through Deep Learning and Transformers. Instead of humans drawing rules, the computer learns them automatically.
The Key Insight: The change wasn't a sudden explosion that wiped out the old ways. It was a slow, evolutionary shift where new ideas grew on top of the old ones.

4. The "Hall of Fame" (Most Influential Books)

The authors identified the top 10 most cited papers (the "Hall of Fame").

The Legend: The book by Dalal and Triggs (2005) is the undisputed king. It is cited more than twice as often as the next most popular book. It laid the foundation for the entire field.
The New Guard: After 2014, the "Hall of Fame" filled up with papers about CNNs (a type of AI) and Transformers (a newer, more powerful AI).
The Warning: Because everyone keeps citing the same few "classic" books, it's hard for brand-new, weird, or different ideas to get noticed. It's like a radio station that only plays the same 10 hits; new artists struggle to get airtime.

5. What Are They Talking About Now? (Keyword Evolution)

If you look at the "tags" on the books, the conversation has changed completely:

Then (1990s): The tags were "Algorithms," "Image Analysis," and "Intelligent Vehicles." The focus was purely on the math.
Now (2020s): The tags are "Pedestrian Safety," "Autonomous Driving," and "Real-world Deployment."
The Big Change: The focus has shifted from "How do we make the math work?" to "How do we keep people safe in the real world?" The conversation is now about safety and systems, not just code.

The Bottom Line

This paper argues that we shouldn't just look at how well a computer detects a person (the technical score). Instead, we should look at the history of the field to understand how it got here.

The authors conclude that AI evolution is like a growing tree, not a series of sudden replacements. The old roots (hand-crafted rules) are still there, supporting the new branches (Deep Learning). However, the field is currently dominated by a few big countries and a few "super-cited" papers, which might be blocking fresh, new ideas from growing. To move forward, we need to connect more researchers globally and focus less on just getting a high score and more on building safe, real-world systems.

Technical Summary: Knowledge Evolution and AI Paradigm Shifts in Pedestrian Detection

Problem Statement
While pedestrian detection is a cornerstone of computer vision and intelligent transportation systems (ITS), existing literature predominantly focuses on algorithmic performance, dataset benchmarks, or specific application challenges. These technical reviews largely overlook the structural evolution of the field's knowledge base and the mechanisms through which Artificial Intelligence (AI) paradigms emerge and transform. There is a significant gap in understanding the macro-level knowledge structure, research networks, and the dynamic transition from traditional methods to modern deep learning and transformer-based approaches. This study addresses the need for a systematic, theory-oriented review that bridges bibliometric analysis with AI paradigm shifts to provide a system-level understanding of how the field has evolved.

Methodology
The study employs a unified bibliometric and scientometric framework to analyze 5,926 publications retrieved from the Scopus database, covering the period from 1990 to 2026. The data collection followed a PRISMA-compatible process, filtering for English-language scientific papers and conference proceedings using specific search queries related to "pedestrian detection," "human detection," and deep learning techniques.

The analytical pipeline consists of three main stages:

Data Preprocessing: Cleaning and normalizing author names, affiliations, and keywords using Python.
Network Analysis: Utilizing VOSviewer for clustering based on modularity optimization and Pajek for main path analysis.
- Collaboration Networks: Analyzed at country, institution, and author levels using metrics such as Total Publications (TP), Total Global Citations (TGC), Total Local Citations (TLC), Degree, Total Link Strength (TLS), and Weighted Betweenness Centrality (WBC).
- Co-citation Analysis: Mapped the intellectual foundations by identifying clusters of frequently co-cited references to reveal knowledge communities.
- Main Path Analysis: Applied global standard and backward local main path algorithms to trace the chronological flow of knowledge and identify the most influential citation pathways.
- Keyword Co-occurrence: Examined the evolution of research topics across three distinct periods (1990–2005, 2006–2015, 2016–2026) to track paradigm shifts.

Key Contributions
The paper makes three primary contributions to the field:

Methodological Framework: It establishes a novel framework integrating bibliometric analysis with AI-oriented theoretical interpretation, moving beyond descriptive statistics to reconstruct the knowledge structure and link it explicitly to the evolution of learning paradigms.
Structural Insights: It provides deep insights into the role of "knowledge hubs" (central nodes), the formation of research clusters, and the mechanisms of knowledge propagation, revealing that the field follows a scale-free and hierarchical structure.
Paradigm Transition Mapping: It conceptualizes the evolution of pedestrian detection not as an abrupt replacement of methods, but as a dynamic process involving intermediate phases of knowledge integration and paradigm coexistence.

Results

Publication Trends: The field exhibits exponential growth ( $R^2 = 0.701$ ), with a dramatic surge in publications post-2010, peaking at over 500 papers annually in recent years. This growth correlates strongly with the adoption of deep learning.
Collaboration Networks: The research landscape is geographically unbalanced. China dominates in publication volume (TP = 2,665) and citations, acting as a central hub. However, countries like the UK, France, and the US demonstrate higher international collaboration intensity (TLS/TP ratio). The network structure is fragmented, with research groups often concentrated within single institutions or regions, limiting cross-cluster knowledge exchange.
Knowledge Evolution (Co-citation & Main Path):
- Green Cluster (Foundational): Represents classical handcrafted feature methods (e.g., HOG, Haar features) anchored by works like Dalal and Triggs [32] and Dollár et al. [6].
- Yellow Cluster (Transitional): Acts as a bridge, containing studies that integrate classical methods with early deep learning or address specific challenges like occlusion (e.g., Hwang et al. [26], Benenson et al. [27]).
- Red Cluster (Modern): Represents the deep learning era, dominated by CNN architectures (R-CNN, YOLO) and, more recently, Transformer-based models.
- Main Path: The global main path confirms a sequential evolution from manual feature design to HOG/SVM, followed by a transition to CNN-based detectors, and finally to attention mechanisms and transformers. The backward local path highlights that traditional methods continued to evolve alongside early deep learning attempts before the latter became dominant.
Keyword Evolution: Early research (1990–2005) focused on "computer vision" and "object recognition." The 2006–2015 period introduced "machine learning" and "ADAS." The 2016–2026 period is characterized by "deep learning," "convolutional neural networks," "pedestrian safety," and "multi-sensor" integration, indicating a shift from pure algorithmic accuracy to system-level safety and application.

Significance and Claims
The authors claim that this study offers a "system-level understanding" of AI field evolution, distinguishing it from traditional reviews that focus solely on technical comparisons. Theoretically, the work posits that AI evolution is a process of knowledge diffusion, accumulation, and restructuring rather than isolated technological advances. It highlights that paradigm shifts occur through coexistence and negotiation between old and new knowledge, rather than immediate replacement.

Practically, the findings provide strategic insights for designing intelligent systems, particularly for safety-critical applications like autonomous driving and Advanced Driver Assistance Systems (ADAS). By identifying "knowledge anchors" and transitional phases, the study aids in selecting appropriate models and architectures. The authors emphasize a necessary shift in perspective from an "algorithm-centric" mindset to a "systems and society" approach, where pedestrian detection is evaluated within the broader context of intelligent transportation, safety, and ethical considerations. The study concludes that while the field has matured technically, future progress requires enhanced international collaboration, interdisciplinary integration, and a focus on real-world deployment challenges such as robustness in adverse conditions and ethical data usage.