Amid mounting pressure on Indian roads, railways, and utilities from ageing infrastructure and manual inspection bottlenecks, AI-driven 3D inspection solutions are stepping in. In an exclusive interview, Ashutosh Bhatnagar of C3D Vision Systems, discusses deep-tech innovations, predictive analytics, immersive visualisation, and the future of automated infrastructure monitoring with EFY’s Akanksha Sondhi Gaur.
Q. Can you briefly outline your company’s profile?
A. Our headquarters are in Mumbai, and we established the business in 2020. We are a private limited firm with three founders: Ashutosh Bhatnagar, Nishchay Malik, and Meenakshi. Nishchay Malik serves as Chief Solution Architect, leading the technical side, while Meenakshi specialises in operations, finance, and HR, and we lead the company’s business development. Our projected turnover for the current year is about 130 million, compared to 68 million last year. We currently have around 30 employees.
Q. Can you briefly explain the innovation and the reason behind the company’s name?
A. We build advanced systems that enable users to observe the world at fine scale and high speed. As technologies evolve rapidly, we aim to help users perceive an additional dimension in their environment: the third dimension of vision. We believe vision is incomplete without depth. While most conventional systems operate in 2D, the real world has depth and dynamism that can only be fully understood in 3D. Our focus is on visualising, acquiring, and processing data in three dimensions. The name ‘C3D Vision’ reflects our mission: enabling comprehensive 3D systems that capture the world as it truly is.
Q. What does your design team look like, and where are they located?
A. Our design team is based in Mumbai, with a few members working from Chennai, Bengaluru, and Delhi. In total, we have around 20 people covering electronics, instrumentation, mechanical engineering, computer vision, and AI. We also include five to six civil engineers and road experts. When we lacked certain capabilities in-house, we conducted our own studies or brought in external experts in specific fields.
For example, designing IP67 enclosures capable of withstanding vehicles covering up to 150,000 km per year on Indian roads required ruggedising the internal electronics and computing systems. Our team addressed most challenges internally, consulting external specialists where needed to build the complete system.
Q. Were the external experts from industry or academia, and were there formal tie-ups?
A. We did not have any academic ties. When we started just after COVID, meeting people in person was difficult due to restrictions. Most of the work was done by our in-house team, and we brought in external experts only when needed. For regulatory aspects, we consulted an expert who advises the automotive and defence industries. We identified specialists within our network for specific challenges and worked with them directly.
Q. Is this tech or business innovation, and what challenges were solved?
A. It is clearly a technology innovation rather than a business innovation. The challenge was building a system that integrates multiple domains, including automotive, civil, optics, mechanical design, geodetics, computer vision, AI, and infrastructure expertise across railways, roads, pipelines, and cables.
It is a mathematically and programming-driven deep-tech product built entirely in-house. This innovation enables new capabilities in managing linear assets.
Q. Where do manual inspections fail in repeatability and accuracy, and how do automated systems fix this?
A. Manual inspections fail in coverage, repeatability, accuracy, and consistency. It is impossible to inspect 100% of roads or survey hundreds of thousands of kilometres multiple times a year. Results vary by inspector and location.
Earlier, GPS capture was manual and required corrections, reducing reliability. Manual methods also lack detail; measuring every crack, pothole, depth, or volume is impractical. Large-scale documentation of road markings, signboards, and assets is nearly impossible.
Automated systems capture continuous, high-accuracy GPS data and process hundreds of kilometres quickly. They compute precise metrics, including average, maximum, and minimum depths and volumes, standardise data, and enable consistent, large-scale monitoring, leading to measurable long-term asset improvement.
Q. Which breakthrough in computer vision or sensing most influenced your product architecture?
A. Several breakthroughs influenced our architecture. The availability of very high-speed cameras and lasers that could be deployed in the right configuration was critical, enabling us to build a highly effective data acquisition system. The ability to synchronise this data with GPS and inertial navigation systems was another major factor.
On the analytics side, recent AI models, combined with our enhancements, enabled the development of this breakthrough product.
Q. What was the key design challenge? How was it solved? Was it in-house or outsourced?
A. We faced multiple design challenges, many with little prior reference. One major challenge was accurately measuring the geolocation of road assets from a moving vehicle. Cameras are mounted at 30–40 degrees, making it complex to calculate precise positions such as median width, median height, curve height, or median opening distance due to perspective variations.
Fusing laser profiler data with RGB data at high speeds while eliminating vibrations and frequency interference required advanced mathematical modelling. Our team studied over 400 research papers to develop the system. All of this work was carried out in-house.
Q. What is the electronics aspect and core technology behind it?
A. This is a technology innovation with a strong electronics and sensor foundation. Our data acquisition system collects inputs from two to eight cameras, laser profilers, IR sensors, and micro- and macrotexture sensors, totalling 17–18 channels.
Designing a system to accurately trigger, acquire, and record this high-speed data—often at speeds exceeding the vehicle’s movement—was highly complex. Power management, enclosure design, laser control through common drivers, and handling multiple data types via USB, RS232, GPS, and proprietary interfaces added to the challenge.
Selecting the right network and data acquisition cards was critical. So far, we have sold four units—three delivered and one in progress.
Q. What key technologies power your inspection stack?
A. Our inspection stack integrates high-speed, high-resolution computer vision cameras with geodetic positioning systems for accurate location tracking. It uses precision optics, including high-clarity lenses and aligned cameras and lasers, to maintain accuracy despite vehicle vibrations.
Advanced electronics support 3D laser profilers and 360° cameras, generating gigabytes of data per second. Automotive and mechanical engineering ensure reliable performance through vehicle dynamics management, finite element analysis, and vibration analysis of mounts and enclosures.
Data from cameras, GPS, INS, encoders, and timing systems is synchronised and unified in a single pipeline. The platform supports one-click export, built-in validation, and automated structuring, making it accessible to civil engineering users without requiring deep technical expertise.
Q. How important is 3D laser profiling vs AI for measurement-grade accuracy?
A. 3D laser profiling is essential for measurement-grade accuracy; AI alone cannot overcome poor data quality. Camera-only systems (dashcams, machine vision, 360°) provide low-resolution outputs and often deliver metre-level accuracy, making crack width, length, or early fault detection unreliable.
They detect only major defects and are suited to basic visual reporting, not maintenance planning. Laser profilers capture sub-millimetre, high-precision 3D data, enabling accurate geometry and depth measurement, such as rut or crack depth. AI supports classification and analysis, but measurement accuracy ultimately depends on high-quality laser data.
Q. How do optics, mechanics, and imaging enable sub-mm crack detection at traffic speeds?
A. Sub-millimetre crack detection requires precise optical, mechanical, and system design. We selected high-quality lenses and cameras with minimal chromatic and barrel aberrations, testing multiple configurations to optimise laser profiling accuracy and account for lens deterioration over time.
Laser wavelength, power, and thermal management were carefully engineered to maintain dimensional stability under varying road, weather, and temperature conditions. Accurate GPS and IMUs ensure precise localisation.
On the software side, architecture decisions—including model selection and cloud versus on-premises deployment—were critical, given the large data volumes (1–4 TB every 1–2 days), making bandwidth and data handling key constraints.
Q. How does sensor fusion ensure spatial accuracy and repeatability?
A. Sensor fusion is critical given the conditions we observe. Ground conditions can change rapidly, whether after rain or within short timeframes.
Without sensor fusion, it is difficult to determine precise data location and timing. For the same location, if we analyse behaviour over time for construction, repair, or maintenance, we require reliable reference points.
Sensor fusion provides consistent, high-accuracy locational references, enabling reliable longitudinal analysis.
Q. How does your sensing pipeline achieve sub-millimetre accuracy?
A. The system is built on proprietary hardware design and algorithms, with precisely aligned cameras and lasers enabling high-speed capture of up to 100 km/h while maintaining ±1 mm accuracy and 0.1 mm Z-resolution.
All optics, lasers, cameras, power, and thermal systems are custom-designed to perform reliably in extreme Indian conditions of up to 50°C. Data is validated onboard and then processed through multiple AI models that reconstruct a full 3D road profile and detect and classify cracks, potholes, rutting, and other defects, measuring their length, width, and depth.
Results are stored in a database and visualised on geo-referenced maps. Sub-millimetre accuracy at high speeds is achieved through tight multi-sensor synchronisation across 8–9 sensors, including high-rate GPS of up to 100 Hz, IMU, lasers, distance encoders, and a hardware clock. Motion-correction algorithms compensate for vehicle dynamics, ensuring accurate real-world 3D reconstruction.
Q. How do you fuse laser depth with RGB while maintaining geometric and visual consistency?
A. We have developed proprietary algorithms and IP to synchronise data from laser profilers and RGB cameras, including a custom-built 360-degree camera system that is time-synchronised with laser profiler data.
The process is complex because RGB cameras are mounted at different heights, positions, and viewing angles, while laser profilers are positioned separately, creating spatial and temporal offsets. Timing jitter can also cause trigger misalignment between sensors.
To address this, we developed synchronisation algorithms that align laser depth profiles with RGB imagery, ensuring consistent geometric and visual representation across varied road surfaces and operating conditions.
Q. How do you approach calibration, validation, and quality control for sub-millimetre inspection accuracy?
A. We have developed a proprietary calibration pipeline using calibration objects machined to an accuracy of 0.01 mm (10 microns) as a reference.
Using these objects, we take reference measurements and compare them with system outputs to calibrate performance. Laser profilers are calibrated across multiple parameters, including positioning, width, depth, and timing.
We follow multiple calibration procedures during system construction. Full system calibration typically takes two to three weeks, including camera synchronisation and accuracy validation.
Q. Which AI/ML models drive defect detection, and how do they prevent performance drift?
A. We use multiple AI models, including object detection, segmentation, and semantic segmentation, which are continuously updated with new data. We have also incorporated CNN-based models, with separate models for detection, classification, and localisation operating together.
Adaptation across infrastructure depends more on road type and condition than geography. After data collection, experts in civil engineering, computer science, and AI select suitable models. The system analyses road construction types, such as asphalt, concrete, and paver blocks, along with condition, and automatically applies the appropriate model to ensure consistent performance.
Q. How do AI models stay robust across varied lighting, dust, repairs, and ageing roads?
A. AI robustness begins with rugged system design to ensure clean and reliable data input. Our systems are IP65/IP67 rated, making them dustproof and waterproof, and capable of operating in conditions ranging from dust storms in Rajasthan to heavy rainfall in the North East, Kerala, and Ratnagiri.
They can operate in severe rain with visibility as low as 2–3 metres. Protection against water ingress and dust contamination, which is critical for high-power lasers, is integrated into the enclosure design.
Built on years of pan-India survey experience, the systems operate within a temperature range of –20°C to 55°C, ensuring consistent data quality for accurate AI detection across diverse and harsh conditions.
Q. How do you prevent AI drift while balancing edge and cloud processing in large-scale 3D inspections?
A. To prevent model drift, we apply five levels of data quality checks before any training or retraining. Continuous field feedback from deployed systems, along with reference surveys, validates outputs and enables corrective updates when required.
Only fully tested and validated models are deployed on edge systems after in-house reliability checks. Due to the large size of laser profiling data, cloud dependence is limited, while inventory data is processed on bare-metal cloud systems for greater control over memory and model optimisation.
The entire pipeline is optimised to handle large-scale 3D data efficiently without workflow bottlenecks.
Q. How do you manage large 3D data and integrate it with GIS systems?
A. Our systems are designed with configurable reporting pipelines, allowing users to define outputs and reports based on their requirements. These outputs can be integrated with asset management systems such as SDM4.
The platform supports multiple standards, including Indian IRC, ASTM, and World Bank SDM4, and allows configuration either before or after processing. This enables different outputs to be generated from the same dataset based on defined thresholds and requirements.
The reporting framework also supports custom formats, enabling seamless integration with GIS, RAM systems, and maintenance workflows.
Q. What is your AI-based Road Inventory System and its key benefits?
A. Our AI-based Road Inventory Management System uses RGB camera data to automate the identification, classification, and geolocation of road assets such as signboards, markings, medians, and gantries.
Traditionally, this process requires large teams manually reviewing video footage. Our system automates the entire workflow, significantly reducing time and effort while improving accuracy. It is fully integrated with our inspection platform and has already processed thousands of kilometres of road data.
Q. What sets Optivision Pro 3D NSV and C3D Leisure Vision apart from other 3D vision or lidar solutions?
A. Our key differentiator is the use of triangulation rather than lidar or conventional 3D vision. This enables sub-millimetre resolution of less than 1 mm and approximately 0.1 mm accuracy across X, Y, and Z axes.
Typical lidar systems operate at 5–15 mm resolution, which can lead to misclassification of road conditions. For example, rutting classifications ranging from 0–5 mm, 5–10 mm, and above 10 mm require high precision.
Our systems provide highly reliable, decision-grade data while supporting high-speed operation, making accuracy and speed the two primary differentiators.
Q. Which core IP pillars will define your long-term technical edge?
A. Our focus areas include predictive analytics, immersive visualisation, and automation. We use large-scale data to build deterioration models based on environmental and infrastructure conditions.
For linear assets such as roads, data is synchronised at intervals of 5–10 metres and presented in a 3D driver’s view, showing faults alongside contextual elements such as signboards and markings. This improves accuracy and confidence in decision-making.
With increasing computing power, workflows are expected to become more autonomous, enabling decision-makers to directly query systems for maintenance planning, simulations, and actionable insights.
Q. What’s your effective go-to-market strategy, and how can AI-generated infrastructure data improve transparency and reduce rework?
A. India has made significant progress in road inspection, particularly for national highways. The National Highways Authority of India has developed robust processes and monitoring systems in recent years.
We follow a direct B2B go-to-market strategy, supported by webinars, business networks, and international engagement. AI-driven infrastructure data enhances transparency, reduces rework, and provides objective insights, supporting more informed decision-making.
Q. Can you share a deployment that showed unexpected ROI and long-term value?
A. One deployment involved a large-scale project spanning over 600 km, where the customer faced challenges due to diverse road types and extreme temperatures. Imported systems struggled to maintain performance, reducing productivity and extending timelines.
Our system enabled the survey to be completed within four to five days, delivering significant time and cost savings. This demonstrated clear ROI through improved efficiency and reliability.
Q. What are the key details of your manufacturing setup, facilities, equipment, and production process?
A. Our manufacturing is based in Mumbai and follows a hybrid model. Vehicle-related work is outsourced to vendors, while final assembly, calibration, and system integration are conducted in-house.
We maintain a cleanroom environment with less than 5 ppm of PM2.5 particles for handling cameras and lenses. The facility includes calibration setups, measurement systems, test rigs, and capabilities for PCB design, including power systems and signal conditioning boards.
We also have infrastructure for optics calibration, dimensional verification, and simulation. While the current setup is compact, it is designed to scale with future growth.
Q. What are the main challenges you are facing in scaling the business rapidly?
A. The primary challenge is building and retaining the right team. The work involves developing new concepts with limited existing references, requiring strong research capabilities.
Finding skilled professionals, managing attrition, and ensuring alignment with the required mindset remain ongoing challenges. As a bootstrapped organisation, careful financial management is also critical.
Q. Do you have any plans for future growth?
A. We are developing new products, including railway systems and a mapping system planned for 2026. Additional products related to railway infrastructure and cable inspection are also under development. These initiatives represent our focus on expanding capabilities and addressing broader infrastructure needs.
Q. Are you seeking new partners, and what is the ideal profile?
A. We are seeking channel partners and distributors with strong networks in the roads and railways sectors, particularly those with a presence across major regional hubs in India.
We are also exploring collaborations with academic institutions for research and development. These partnerships aim to enhance solutions in infrastructure management, road safety, urban planning, and accident reduction.
