Sober Guardian

The Project

The problem of driving under the influence continues to cause an alarming number of car accidents and crashes, which in the US claims over 10,000 lives and costs about $194 billion annually. Currently, there is no effective approach to detecting impairment while someone is driving, except for stopping them and conducting a sobriety test. To address this, I have developed Sober Guardian, an AI system and a mobile app capable of instantly detecting whether a driver is under the influence of substances such as alcohol, marijuana, cocaine, and others, and automatically sending notifications. This novel and innovative system utilizes computer vision-based AI and machine learning technologies, including facial recognition and visual analysis, to provide real-time identification of impaired drivers by comparing the driver's features to a pre-trained dataset. If the system detects any signs of impairment, it produces a prompt notification, automatically sending warnings to the driver. The driver will then have the option to notify a preset list of contacts or connect with a ride sharing service and share their location. The AI model was trained using a small initial dataset, which underwent preprocessing (photo splitting and organization) and exploratory analysis (recognition of patterns). The training process was carried out using Google Collaboratory, a Convolution Neural Network, and the TensorFlow Lite model, resulting in an accurate prediction of the driver's sobriety status. This is then applied to a predicting model that can be utilized by various departments, including law enforcement, for real-time detection of substance influence. The model’s accuracy was optimized by adjusting the batch size and epochs during testing (60 epochs, 32 batches, and 0.001 learning rate). The result was a high accuracy rate for detecting drunk drivers (96%) and sober drivers (88%) which exceeded my expectation of 85% overall accuracy. This improved model was then successfully applied to an external test set, utilizing image and video. It’s also an intelligent and practical app designed with a sophisticated front-end and back-end structure and architecture, integrating TensorFlow lite model. The coding languages I used are Python and FlutterFlow. The tools I used for this project include: Software: - Python’s TensorFlow - Python’s Keras - Google Colabatory - Adobe Premiere Pro - PyTorch - Python and Python Flask - RPi.GPIO - Flutter Flow and Dart - Android Studio Cloud and Database: - Google Firebase Database My project, Sober Guardian, has shown that using AI facial recognition and visual analysis can effectively detect and prevent driving under the influence. My ultimate goal is to integrate these technologies into vehicles and smart devices to ensure road safety and reduce accidents and save lives caused by drunk driving.

View this project on YouTube

Team Comments

In 2017, I lost my close friend to a DUI incident. Every 15 minutes, a teenager dies due to drunk driving. Every year in the U.S., driving under influence claims about 10,000 lives and costs approximately $194 billion. Currently, there is no effective approach to identify DUI drivers.

It was difficult to acquire high-quality images of human subjects before and after alcohol consumption. I used a small sample pool of images collected from 53 participants, then used Google Collaboratory and cv2, a Python package to cut each image into 4 separate images, to create a larger dataset.

I would create an embedded physical device using the Internet-Of-Things (IoT)-based air sensor system, to monitor in-car environment, integrate the AI model with the mobile app, and offer it as a safety option in new vehicles, to prevent driving under the influence accidents and save many lives.