Similarity-based Frame Screening for Edge-Cloud In-Cabin Monitoring

In-Cabin Monitoring systems are essential for autonomous driving, enabling accurate assessment of driver and passenger conditions during handovers. Vision-Language Models (VLMs) are effective for this task but face challenges due to high computational demands and limited communication bandwidth. To address these issues, we propose a lightweight edge-processing technique that screens frames before cloud-based VLM analysis. Our method computes visual feature similarity to distinguish normal driving from abnormal behaviors and transmits only abnormal frames for detailed VLM analysis. Evaluation results show that this approach reduces communication volume by approximately 99% while maintaining low processing costs on the edge device.