AI Security Cameras are Ditching the Cloud. It's Now All About Edge Inference

AI is changing the surveillance industry in more ways than one. While the first wave of AI-powered video analytics relied heavily on cloud computing, the next phase is shifting intelligence directly into cameras, enabling devices to analyse events locally.

The transition is being driven not only by the need for faster decision-making but also by India's growing focus on cybersecurity, trusted hardware, and indigenous semiconductor development.

Government regulations have also accelerated the shift. Over the past two years, India has tightened security norms for internet-connected CCTV cameras, making compliance with Essential Requirements (ERs) mandatory. The framework requires manufacturers to disclose the origin of security-critical components such as system-on-chips (SoCs), undergo cybersecurity testing, and comply with Standardisation Testing and Quality Certification (STQC) norms before deploying products in sensitive environments. Government departments have also been directed to procure only compliant CCTV equipment.

Vendors are also now increasingly embedding AI directly into camera hardware. Running AI locally requires more than moving an existing video analytics model onto a camera. The hardware needs to support AI inference alongside video encoding, image processing, and memory workloads within the power and thermal constraints of an edge device.

Singapore-based fabless semiconductor startup StraitsMicro’s proposed architecture illustrates this convergence. Its Visora family combines a multicore 64-bit CPU with hardware-accelerated AI, low-power DDR5, H.265 encoding, and an imaging pipeline designed to reduce memory-bandwidth requirements. Depending on the variant, the platform supports workloads ranging from human and motion detection to crowd and behaviour analytics. 

"The future of surveillance lies in distributed intelligence, where the camera evolves from a vision device to an intelligent sensor capable of understanding its environment," Sanjeev Sehgal, Founder and Managing Director of Sparsh CCTV, tells AIM.

AI-enabled cameras can detect perimeter intrusions, unattended baggage, crowd build-up, fire, or traffic violations in real-time. They also simultaneously send organisations alerts while transmitting only relevant events or metadata instead of raw video streams. This significantly reduces latency while lowering bandwidth consumption and storage requirements.

States and government bodies are also increasingly deploying AI cameras for smart enforcement systems. Goa will implement 55 AI traffic cameras by January next year, while the National Highways Authority of India plans to deploy AI-powered Dashcam Analytics Service across 40,000 km of national highways to automate road maintenance.

Sparsh CCTV is deploying its AI cameras in projects including the USBRL railway corridor, the Chenab Bridge, NHAI's multi-lane free-flow tolling programme, the Jammu Railway Division, and Prayagraj Maha Kumbh.

The Economics

The economics are also changing. While AI-enabled cameras carry a higher upfront cost because they include dedicated AI processors, vendors argue that enterprises are increasingly evaluating surveillance deployments based on total cost of ownership rather than the price of individual cameras.

Local inference reduces cloud storage, lowers server workloads, and decreases network infrastructure costs, particularly in deployments involving thousands of cameras across airports, railway stations, highways, factories, and smart cities. As per an analysis by Harry Hua, Technical Director at AI and IoT vision brand CamThink, edge AI in cameras can reduce costs by 67% while enabling sub-second local response.

"The objective is not to choose between hardware AI and cloud AI but to build the right AI architecture for the application," Sehgal says.

As per Market Research Future analysis, the AI CCTV industry in India is projected to grow from $1.8 billion in 2025 to $7.6 billion by 2035.

Hybrid Architecture

Industry observers believe surveillance is gradually adopting a hybrid architecture. Edge AI is increasingly handling time-sensitive operational decisions, while cloud platforms continue to provide historical investigations, enterprise-wide analytics, fleet management, and long-term storage.

“By running AI models close to the camera, only meaningful events and alerts need to be transmitted to command centres. This reduces network traffic, enables near real-time responses, improves reliability, and allows critical decisions to be made even when network connectivity is limited,” Hari Kishan, CTO of Bengaluru-based AI solutions company Pancait Smart Systems, wrote in a LinkedIn post.

Another factor accelerating edge AI adoption is cybersecurity. Analysing video directly on the device reduces the amount of sensitive footage moving across networks, allowing organisations to transmit only alerts, or event-based recordings where necessary. For sectors such as defence, aviation, railways, and critical infrastructure, this supports broader data sovereignty and cyber resilience objectives.

However, vendors note that secure surveillance depends on more than where AI is processed. Secure firmware, encrypted communications, authenticated software updates, and trusted hardware remain equally important components of a resilient surveillance architecture. 

Human verification also needs to stay in the loop, as error rates remain volatile. Around 90–95% of violations flagged by Greece’s new AI-powered traffic cameras in June in a pilot programme were reportedly found to be incorrect after police verification. Meanwhile, the police department in the city of Roseville, US, found a 71% failure rate with their AI traffic cameras, which would mistake one licence plate number for another.

“There is no single person or organisation that automatically audits bias and false alerts at city scale. It has to be a shared responsibility between the AI provider, the system operator/deployer, independent evaluators, and—depending on the use case—regulators or government oversight bodies,” believes Kishan.

The emergence of new processors reflects a broader shift in AI infrastructure. As models become smaller and more efficient, inference is increasingly moving from centralised cloud servers to specialised edge hardware that delivers real-time performance.