AI Camera Platform for Exterior Vehicle Access

Project in a Nutshell: 

Promwad ran an internal R&D project to build a reusable AI camera platform for exterior automotive access. The goal: validate that gesture recognition, driver face identification, and parking surveillance can run on a single automotive-grade chip, within the latency and power constraints production requires, with ASPICE CL2 process readiness considered from the architecture stage. 

The platform delivers: gesture decision under 120ms, face recognition accuracy above 99%, deep sleep power under 100µA. Estimated hardware cost under $80 at 10K volume, under $49 at 1M+. All processing runs on-device, no cloud dependency. 

Challenge

Biometric and gesture-based vehicle access systems have already entered mass production at many automakers in Asia and Europe; these include keyless entry systems with facial recognition and gesture-controlled door opening. Demand for such systems has taken hold. But most OEMs outside that early group are still stuck in R&D, blocked by the same three issues: 

• Off-the-shelf computer vision models don't scale to automotive processors without significant custom work. 

• Power budgets for always-on exterior cameras are tight enough that a wrong architectural choice can force a hardware redesign. 

• ASPICE process requirements cannot be tacked on later; they must be integrated from the start.

Solution

Promwad built the platform to solve all three problems from a single hardware foundation.

The platform integrates gesture recognition, driver identification, and parking surveillance on a single Ambarella CVflow System-on-Chip. Mode switching is handled in firmware, so the same hardware works across different vehicle programs. 

We chose the CVflow family because it runs image processing and AI inference on one automotive-grade chip suited to edge AI in automotive camera applications. The development process was aligned with ASPICE CL2 requirements from the start.

Gesture-based access. The prototype captures the first video frame within 100 ms after the CAN wake-up signal and returns a gesture decision within 120 ms, measured under the tested configuration. This is fast enough for everything to happen instantly at the door. The pipeline uses the Google MediaPipe framework, adapted to run on our hardware. 

Facial profiling. Recognition begins as the driver approaches the vehicle. Once a match is confirmed, their personalized profile is loaded via the vehicle’s bus before they even sit down: seat position, mirrors, climate control, and driving mode. In a shared-use vehicle, this eliminates the need for manual adjustments every time the driver changes. In the prototype test setup, face verification accuracy exceeded 99% for two enrolled users under defined test conditions. Recognition runs entirely on-device, no biometric data leaves the vehicle. 

Sentry Vision. While parked, the system sleeps at under 100µA. A motion trigger wakes it, the system classifies what it sees (someone walking past vs. inspecting the car), pushes a notification to the owner's app, and saves a short clip. No continuous recording reduces storage volume and supports privacy-by-design principles. Final GDPR compliance depends on retention policy, user consent, biometric data handling, and OEM deployment context. 

Results

For automakers and Tier 1 suppliers launching development programs for external vision systems, this platform eliminates the configuration phase that typically delays production readiness. MediaPipe models have already been ported to CVflow, the system architecture has been validated, and the ASPICE CL2 process groundwork is already in place. The development team starts with a ready-to-use base version backed by proven test results, rather than starting from scratch. 

The architecture scales within the existing ECU boundaries: additional gestures, more registered users for facial recognition, and integration with vehicle or fleet management applications. None of this requires hardware redesign. 

What We Offer

The platform is available as a starting point for two types of engagements. 

Adapt to your program. If you're building an exterior vision system for a vehicle or industrial product, we can adapt the platform to your hardware, use case, and integration requirements. Model porting, power validation, and ASPICE CL2 documentation groundwork are already done. 

Proof of Concept. A fixed-scope engagement to validate gesture access and face recognition on your target vehicle, with defined success criteria and a clear production integration path at the end. 

More of What We Do for Automotive Vision

• Automotive Vision Systems: explore our expertise in automotive camera systems for passenger cars, trucks, and commercial vehicles.
   
• Access Control System: see how we built and tested an embedded facial recognition system for car-sharing access control.
   
• Exploring Lattice FPGA for Digital Rear-View Mirrors: read our breakdown of why Lattice FPGAs work well for digital rear-view mirrors and camera-based ADAS.