Overview

As part of the IMAP Project, our goal was to create a safer, smarter driving experience by enhancing our Intelligent Behavior Detection module. This activity focused on improving detection speed, accuracy, and integration efficiency — ensuring that real-time driver monitoring can operate seamlessly in next-generation smart dashcams.

 

The Challenge

Traditional face and behavior detection systems faced several limitations:

  • Slower detection speeds on earlier hardware (Raspberry Pi Compute Module 4).
  • Lower detection accuracy using certain deep-learning frameworks (e.g., DLIB, OpenVino).
  • Limited camera compatibility, resulting in reduced field of view and slower responsiveness.
  • High computational demand affecting real-time performance in compact devices.

These issues collectively impacted our ability to deliver fast, reliable, and cost-efficient driver safety features.

 

Our Approach

We re-engineered the program by:

1. Upgrading the Core Hardware
 Transitioning from Raspberry Pi Compute Module 4 to Compute Module 5 offered significant gains in processing power, memory, and stability — all crucial for deep learning inference in real-time.

2. Improving the Visual Pipeline

  • Switched to Camera Module 3 (SONY IMX708) for a wider field of view, better resolution, low-light performance, HDR support, and autofocus.
  • Enhanced facial coverage and motion tracking for improved safety event detection.

3. Optimizing AI Model Execution

  • Migrated face detection models to ONNX Runtime, using the ultra-lightweight RFB-320 model.
  • Achieved faster inference with better hardware acceleration on Raspberry Pi CM5.
  • Retained eye detection through OpenVino for accuracy in detecting drowsiness-related cues.

4. Enhancing User Alerts
 Introduced a buzzer-based alert system to provide immediate audio feedback when unsafe driving behavior is detected.

5. Preparing for Future Expansion

  • Explored integration with GPS and accelerometer data for speed and motion-based safety context.
  • Maintained minimal GPIO usage to allow compatibility with other smart dashcam features.

 

The Results

  • Performance Improvement: Faster detection speeds and improved accuracy across key driver behaviors including calling, drowsiness, and visual distraction.
  • Real-Time Readiness: Achieved low-latency detection suitable for compact, in-vehicle systems.
  • Integration-Friendly Design: Created a robust yet scalable solution that can be deployed in next-generation smart dashcams without significant cost or size trade-offs.

 

Impact

By optimizing the Intelligent Behavior Detection program, we moved closer to our mission of safer roads through technology. This advancement not only enhances driver awareness and response but also supports insurance, fleet, and consumer automotive partners in deploying smarter, more proactive safety solutions.