Vision Based Driver Smoking Behavior Detection Using Surveillance Camera Images

Intelligent traffic enforcement has gained immense traction in the computer vision community. Recently, automated violation enforcement methods have been proposed towards seat belt violation, cell phone violation and occupancy violation detection tasks. Smoking while driving is another common violation type that has been prohibited in many countries. Smoking inspections are typically performed manually by the road side officers. In this study, we propose an automated approach towards driver smoking behavior detection using near infrared (NIR) surveillance camera images. During the puff, cigarette tip reaches 800-900 degrees C creating a hot-spot on the NIR image. Proposed method aims to detect these hot-spots around the drivers' head region. First, we utilize a deep learning based object detection technique to localize the front windshield and driver head region, sequentially. Next, we perform a dual window (local) anomaly detector on the localized region to determine white hot-spot, hence, the driver smoking behavior. We have collected 1472 real world NIR images to evaluate the performance of the proposed approach. Proposed method achieved an overall accuracy rate of % 84 and sensitivity rate of % 70 on the test set.