Future of Dashcam-Lok

What are some problems the follow-up project to Dashcam-Lok could focus on?

Assuming that the current phase of Dashcam-Lok ends with delivering:

• 50k hours of diverse driving data (RGB, with metadata and possibly unanonymized)
• Model trained for 'person' deployed on the edge-device
• Model trained for the other 'n' classes deployed on the workstation
  And probably delivering:
• The trained model(s) and weight(s)
• The labeled data used for train/val/test
• Statistics on the redundancy - number of classes (and objects) encountered in total

Three approaches
Data - Approach focused on scaling up data collection efficiently
AI - Approach focused on using the collected data for AI research