AI Feasibility Study for Traffic Video Analysis
Can AI turn video from traffic cameras into reliable counts of road users, and what does it take beyond the model? For a client in the mobility sector EMBIQ assessed detection approaches, data protection and usability for non-programmers, and backed the answers with a cloud proof of concept.
Technologies used
Python, OpenCV, YOLO, COCO dataset, time-series database, synthetic training data, cloud GPU instance
Tools
Git, cloud console, dashboards
3
Project members
3 months
Project length
AI Development
Category
Transportation
Branch
The challenge
Our client wanted to know whether video from traffic cameras could be turned into statistics about road users with AI, before investing in development. The study had three questions to answer: which detection approach is fit for counting different types of road users, how to handle data protection and ethics when processing images of public spaces, and whether such a tool could be used by people who do not write code. The practical constraints turned out to be as important as the technical ones. Public traffic cameras are rarely offered as open video streams, so any concept built on them has to start with data access. A demo also had to be something that could be shown without special hardware, and the question of who would operate such a tool in the long run had to be answered alongside the technical ones.
The client did not need another object detector; they needed to know whether the whole idea holds up outside the lab. So we treated it as a feasibility question first: data access, privacy and who will actually operate the tool.
Key Features
- Review of machine-learning and deep-learning approaches for traffic image analysis and choice of a detection approach
- Proof of concept that detects and counts road users in a video stream with a pre-trained open-source detector
- Configurable object classes, so counted categories can change without new code
- Counts stored over time and presented as simple charts
- Cloud setup with cost controls, usable without a local GPU
- Assessment of synthetic training data for classes missing from public datasets
- Data protection and ethics analysis for processing images of public spaces
The result
EMBIQ delivered a feasibility report together with a working cloud proof of concept, documentation and a video walkthrough, so the approach could be demonstrated without installing anything locally. The most useful findings were not about model accuracy. Data access and the operating model of the future user mattered as much as the choice of algorithm. With these answers the client could decide where the concept makes sense and plan the next steps.
- Feasibility report with recommendations and next steps
- Working cloud proof of concept with video walkthrough and user documentation
- Configurable object classes, extendable with further training data
- Cost-controlled cloud environment
- Clear view of non-technical barriers such as data access and operating model
Keeping the proof of concept in the cloud, with one screen to start and stop the analysis, meant it could be shown to anyone the same day. The report then gave honest answers on what would and would not scale.
What kind of competencies are you looking for?
Contact me to discuss your requirements and build your dedicated IT team together with us.
Agnes
business development specialist
Quote a project! Get advice.
Let’s talk about your project
Drop us a line or book a short intro call — we’ll get back to you with the right people on our side.
E-mail usinfo@embiq.comBook an intro call (opens in a new tab)
Let’s investigate your project concept and its current status together.
Expect an initial project scope proposal, time and cost estimation from us.
The consultancy will be protected by the NDA.