The Kolkata crush and mechanics of pandal safety: Why smart city sensors fail in high-density downpours | Explained

For the first time in its history, Kolkata on Saturday witnessed a stampede during Durga Puja celebrations that led to a loss of 19 lives, sending shockwaves across West Bengal.

The incident occurred at the Chorbagan Sarbojonin Durga Puja pandal, and besides the 19 deaths, 25 other people were injured in the incident.

The crisis that unfolded in Kolkata exposes a recurring blind spot at the intersection of urban technology, event infrastructure economics, and liability insurance.

While public chatter is blaming either the police or the organisers of the North Kolkata puja of mismanagement, a look at the underlying technology that remains inherently vulnerable to structural, weather, and crowd movement failures.

The physics of AI failure: Why smart city cameras go blind in surges

Over the past five years, a number of municipal corporations and event management firms across major Indian metro cities have deployed AI-driven video analytics, computer vision (CV), and LiDAR sensors for real-time crowd density estimation. Kolkata Police, however, does not use LiDAR sensors and instead depends on an Integrated Command and Control centre where it gets visual inputs from thousands of CCTV feeds across the city. AI algorithms then process these streams in real time to generate alerts where localised crowd density crosses safety thresholds. It does use AI-driven video analytics and computer vision (CV) for traffic monitoring and crowd management.

However, these systems face severe technical degradation during high-density, seasonal festival surges due to specific algorithmic and hardware constraints:

1. Occlusion and Head-Detection Collapse: Standard algorithms used for crowd-counting rely on Convolutional Neural Networks (CNNs) trained to identify human heads and shoulders. When crowd density crosses a critical threshold, which is typically 4 to 5 people per square meter, individual visual boundaries tend to blur. In heavy rain, with umbrellas, rain ponchos, and dynamic reflection off wet pavement (which gets more enhanced due to the extra lighting during Durga Puja), head-detection accuracy drops by over 65%, as per a research.

2. Re-Identification (Re-ID) Errors: Multi-camera tracking systems use Re-ID feature vectors to monitor flow rate as well as direction across narrow street grids. When the crowd density is high, shadow clutter and low-light conditions cause feature vectors to fail, thus causing software to undercount static bottleneck buildup while also misinterpreting directional flow vectors.

Infrastructure issue

Pandals are classified as temporary lightweight structures under Indian Standard Codes (IS 875 Part 3). However, temporary approvals rarely account for dynamic wet loading—the combined weight of rainwater-soaked fabric, shifting wind pressures, and sideways crowd force. When unexpected downpours occur, the structural dead-weight of non-waterproof fabric can double within minutes, drastically reducing the lateral force required for a perimeter barrier or support column to fail.