Passenger counting sounds simple. Count people getting on, count people getting off. But try doing it accurately in an Indian public bus — 50°C heat, no reliable internet, crowds pushing through a single narrow door simultaneously, pitch darkness at night stops, and constant vibration on potholed roads.
We tried the obvious approaches first. They all failed.
Why Cloud Won’t Work Here
The default architecture for any modern AI system is: capture video, upload to cloud, process, return results. This works great in retail stores with stable WiFi. In Indian transit, it falls apart immediately.
- Connectivity is unreliable. Buses pass through tunnels, rural stretches, and areas with zero signal. You can’t stream video to a server that’s unreachable for 40% of the route.
- Latency kills accuracy. Passenger counting needs frame-by-frame tracking. By the time a cloud server processes a frame, the bus has moved, people have shifted, and the count is wrong.
- Data costs are prohibitive. Streaming even compressed video from thousands of buses 24/7 would cost more than the entire analytics platform.
The Edge AI Approach
We designed a self-contained unit — camera, compute, and storage in a single ruggedized box powered by an NVIDIA Jetson. Everything runs on-device.
The CV model detects and tracks each person as they board and alight. It handles:
- Simultaneous entry and exit through the same door — a pattern unique to Indian buses where passengers push both ways at once
- Crew filtering — the driver and conductor aren’t passengers; the system learns to exclude them
- Zero-light conditions — IR-capable cameras with models trained on nighttime footage
- Crowding — when 15 people board in 8 seconds, the tracker needs to maintain identity across severe occlusion
Results are stored locally and synced to the cloud dashboard whenever connectivity is available. Trip reports include video clips for each boarding event so operators can verify counts.
The Hardware Challenge
Consumer hardware doesn’t survive Indian transit conditions. We learned this the hard way with early prototypes.
- Heat: Interior bus temperatures regularly exceed 50°C. Standard compute boards throttle or shut down. We had to design passive cooling solutions that work without fans (dust kills fans in weeks).
- Vibration: Constant road vibration loosens connectors and damages SD cards. We moved to soldered storage and vibration-dampened mounting.
- Power: Bus electrical systems are unpredictable — voltage spikes, sudden cutoffs, reverse polarity. The unit needs to handle all of it gracefully and resume operation after any power event.
What This Means for Other Industries
The engineering challenge of making AI work in harsh, disconnected environments isn’t unique to transit. Factories, construction sites, agricultural operations, and remote installations all face similar constraints.
If your use case involves:
- Unreliable or no internet connectivity
- Extreme environmental conditions
- Real-time processing requirements
- Cost sensitivity at scale
Then edge AI isn’t just an option — it’s the only viable architecture.
Conclusion
We built this system because the problem demanded it. No off-the-shelf solution handled the combination of Indian boarding patterns, harsh conditions, and connectivity constraints. The result is a passenger counting system that’s been proven in the field — and the engineering lessons apply far beyond buses.
Interested in edge AI for your use case? Talk to our team — we’ve solved the hard problems so you don’t have to.