Iphone + GPT-4 Vision API = Autonomous Security Cam System

An iPhone-based security camera can be turned into an autonomous change-detection system by pairing it with the GPT-4 Vision API, automatically logging what’s different between successive snapshots and emailing an hourly summary. The core workflow is simple: the iPhone captures an image every five minutes, uploads it to Google Drive, and a script pulls the newest frame, compares it against the previous one, and uses GPT-4 Vision to generate a structured description of changes. Those descriptions are then compiled into logs and sent to the user via the Mailgun API.

The system’s logic hinges on image-to-image comparison. Each iteration downloads the latest image from Google Drive and feeds it—along with the prior image—into GPT-4 Vision (using the “gpt-4 vision preview” model). A carefully written prompt instructs the model to look for specific categories of change: time-of-day and weather consistency, counts of cars, bikes, people, and animals, and any “unusual activities.” The prompt also requests a consistent output format, including a timestamp, so the resulting records can be used as a practical audit trail rather than a free-form description.

To turn frequent checks into something actionable, the script runs on a timed loop (a “trup”/cron-like schedule with a 350-second sleep between checks). It collects per-comparison logs for 1 hour by storing 12 reports (12 × 5 minutes ≈ 60 minutes). When the counter reaches 12, the system summarizes the last hour of observations into a single “log hour summary” and emails it. After sending, it clears the stored reports and starts the next hour-long cycle.

The implementation ties together several services: Google Drive for storage and retrieval, Google Drive API for authentication and downloading the latest image, GPT-4 Vision for change detection and narrative logging, and Mailgun for email delivery. The Python functions reflect this modular design: one function handles Drive authentication and image download; another analyzes two images and generates the change description using GPT-4 Vision; a third sends the email using Mailgun; and a GPT chat function produces the final log text and summaries.

In a test run lasting over an hour, the system produced an email while the user was away. One comparison showed no meaningful differences—consistent daytime lighting, stable weather, and the same number of cars, with no bikes, people, animals, or unusual activity detected. A later comparison flagged a clear change: a neighbor appearing in the second frame. The log noted that one person was visible in the later image, described the activity as routine (walking through the yard), and aligned the event with the correct timestamp. The hourly email summarized the broader pattern: mostly static scenes, clear weather, car counts varying within a small range, and people appearing only in a few comparisons.

The results suggest the approach works reliably for basic scene monitoring, though the creator also points to prompt limitations—people were sometimes only partially described, and the neighborhood’s low activity reduced the volume of detected events. Still, the project demonstrates a practical path from consumer hardware (an iPhone) to automated, model-driven security reporting without building a full computer-vision pipeline from scratch.