Automated crowd counting in extreme environments such as the Hajj pilgrimage represents one of the most complex challenges for modern computer vision. Cameras capture the scene from near-vertical angles, people massively occlude each other, and a single frame can contain over a thousand individuals. Traditional person counting models, trained on urban or indoor scenarios, fail dramatically when faced with these conditions. That is why the research community has begun exploring zero-shot paradigms, which do not require specific retraining for each new domain. A recent study on the HAJJv2-CrowdCount dataset, which provides per-second annotations for test videos, evaluated three cutting-edge approaches: YOLO-World (open-vocabulary detector), APGCC (point-based counter), and SAM3Count (promptable segmenter). The results reveal a crucial inversion: while SAM3Count achieves the best overall mean absolute error (MAE 70.4), in the densest frames its performance plummets (MAE exceeding 300), whereas the point-based counter APGCC degrades much more gracefully (MAE 114.9). This inversion has direct implications for Hajj safety management: reliable counts are needed precisely in the most crowded and occluded scenes. From a technical perspective, the study demonstrates that no single architecture is sufficient and that enterprise solutions must combine multiple approaches, dynamically adapting to scene density.
To address this challenge, organizations require custom software systems that integrate AI models, cloud platforms, and Business Intelligence dashboards. For example, a real-time counting system must process streaming video using scalable infrastructure on AWS or Azure, apply detection and tracking algorithms, and visualize data in interactive dashboards such as Power BI. Furthermore, cybersecurity is critical: video data contains sensitive pilgrim information and must be protected from unauthorized access. A company like Q2BSTUDIO, specialized in developing custom software, can build these platforms by combining artificial intelligence, cloud computing, and cybersecurity. Its integrated approach allows not only deploying zero-shot models like those evaluated but also orchestrating AI agents that monitor density and trigger automatic alerts, all on secure cloud infrastructures.
The benchmark also highlights the importance of autonomous AI agents. In the densest frames, where detection- and segmentation-based methods collapse, an intelligent agent could dynamically switch to the point-based approach, maintaining accuracy. This adaptive decision-making capability is precisely the kind of functionality that Q2BSTUDIO implements in its AI projects. The company not only offers pre-trained models but designs complete pipelines integrating computer vision, cloud processing, and BI for real-time informed decisions. For instance, a Power BI dashboard could show density evolution in different areas of Mecca, alerting security teams when critical thresholds are exceeded.
Beyond Hajj, this type of technology has applications in stadiums, concerts, airports, and any mass event. The combination of cloud AWS/Azure for scaling processing, cybersecurity to protect data, and BI/Power BI to visualize information creates a robust ecosystem. Q2BSTUDIO, with its expertise in custom software development, helps companies build these systems, avoiding the limitations of generic commercial solutions. Ultimately, success in crowd counting does not depend solely on the best AI model, but on the overall architecture that supports it: from video capture to operational decision-making.
In summary, the study on zero-shot benchmarking in Hajj reveals that there is no universal solution; the key lies in adaptability and integration of multiple techniques. Companies wishing to implement reliable person counting systems should turn to custom applications that unite AI, cloud, BI, and cybersecurity. Q2BSTUDIO is ready to face that challenge, offering a multidisciplinary team that turns academic research into concrete and effective business solutions.
