AI Crowd Management Fails Again: Puri Stampede Tragedy

AI surveillance has become a popular tool for crowd management in India, especially at large events and pilgrimage sites, aiming to enhance safety and efficiency. However, its success has been inconsistent, as evidenced by recent incidents.

AI-powered systems are designed to analyze crowd density, identify congested areas, and detect unusual activities. These systems use facial recognition to locate missing individuals and behavioral anomaly detection to spot suspicious behavior. Drones and aerial surveillance systems are also deployed to monitor large areas, manage traffic, and track crowd movement.

Tarun Wig, co-founder of Innefu Labs, explains, “Our system continuously learns from baseline crowd behavior. A sudden spike in velocity, erratic motion vectors, or group-level directional shifts trigger real-time alerts for possible panic or stampede scenarios.” Innefu’s systems use convolutional neural networks (CNNs) and Transformer-based vision models for crowd density estimation, while unsupervised learning models like autoencoders and recurrent neural networks (RNNs) track deviations from typical movement patterns.

The deployment of AI and IoT drone systems is increasing for various events, from temples and festivals to airports and major sporting events. However, while AI in crowd surveillance has advanced, there is a pressing need for better governance frameworks, data rights protection, and accuracy standards to safeguard privacy.

Innefu’s surveillance stack is designed with open APIs, allowing seamless integration with third-party drones, thermal sensors, acoustic detectors, and other IoT components in smart city environments. The data is harmonized via their central AI fusion platform for real-time decision-making.

The most recent deployment of similar systems was seen during the Rath Yatra in Ahmedabad on June 27, where an estimated 1.5 million people were expected to attend. Despite the advanced technology, recent stampede incidents, such as the RCB victory celebration in Bengaluru, which caused 11 deaths and injuries, have prompted cities to enhance security measures.

For the first time, Ahmedabad’s security measures included AI-based surveillance systems. According to PTI, the AI system could notify the police and fire department in the event of a fire and alert the police control room about potential overcrowding. Around 14-15 lakh people attended the Rath Yatra.

In Puri, the pilgrim city also saw over 275 AI-enabled CCTV cameras and drones deployed alongside 10,000 security personnel, supported by a real-time WhatsApp chatbot for public assistance and an integrated command and control center for centralized coordination and data-driven decision-making to manage nearly 1.5 million devotees.

An Odisha-based defense tech startup, IG Drones, partnered with the state government for advanced aerial surveillance and counter-drone systems to ensure seamless monitoring and threat neutralization during the event. Bodhisattwa Sanghapriya, founder & CEO of IG Drones, stated, “This year, by integrating our advanced drone surveillance with anti-drone defense systems, we’re helping ensure that the Yatra remains peaceful, secure, and uninterrupted.”

Despite these measures, the government saw another fatal failure in managing the crowd. Three people died and over 50 were injured during the Rath Yatra in Puri on Sunday morning. When asked to comment post the accident, IG Drone’s spokesperson issued a statement saying, “At this moment, we are not in a position to officially comment on any specific matter related to this. IG Drones remains fully committed to Rath Yatra 2025 and has been actively working to provide technology solutions for the same.”

Similarly, the AI cameras that monitored the crowd during the Maha Kumbh in Uttar Pradesh in January this year were a failure. Reports stated that more than 30 people died, and 60 sustained injuries after a stampede broke out at the snana site.

The challenges are numerous: Many tier 2 and 3 cities lack the backend infrastructure (bandwidth, compute power, and trained staff) needed for effective real-time analytics. Inconsistent CCTV quality hampers model performance, and facial recognition often struggles with darker skin tones, crowded settings, and partial obstructions. False positives and negatives can lead to misidentification, especially in sensitive community contexts.

India also lacks a comprehensive law to protect personal data. AI systems often collect identifiable data without consent, raising concerns about potential misuse and the tracking of individuals outside crowded spaces.

As the use of AI in crowd management continues to evolve, it is clear that more robust governance, infrastructure, and ethical frameworks are necessary to ensure the safety and privacy of the public.