Activity Recognition on the Dropcam Timeline
Dropcam stands patiently all day, waiting to catch sight of something interesting. To help you sort through hours of footage, we provide timeline markers and alerts when motion activity occurs. However, many customers receive alerts about things they consider to be “false positives,” like shadows, flickering light or trees moving in the wind. Customers might also receive alerts about activities they already expect, such as people entering a room.
Our goal is to make Dropcam smarter so you can receive alerts about the types of activities you care about. That’s why we’re excited to introduce Activity Recognition, a new beta feature that all Cloud Recording (CVR) subscribers can access through our web app.
How does it work?
Dropcam uses software algorithms to recognize motion patterns in your video and organizes them into categories based on patterns it detects most frequently. Each category is assigned a color and when a new activity is recognized, a color-coded motion marker appears on your timeline. You can:
- Assign names to categories
- Filter what activities appear on your timeline
- Customize what alerts you receive by category
You will start to see your first categories after about a day, and a few more each day until Dropcam learns most of the typical activities in your camera’s view. Categories are dynamic and become more refined over time.
The Benefits of Activity Recognition
To get the most out of Activity Recognition, let’s start with some basics. First off, it’s best to keep your camera in one location so Dropcam can learn about all the activities that happen there and create more useful categories over time. It’s also good to maintain a consistent zoom level and position your camera to capture a large area of the room. As more activities are recognized, you’ll be able to:
- Make alerts more informative: Check out several motion markers that have the same color in your timeline. When you see what pattern is recognized, you can assign a meaningful label and decide what alerts you’d like to receive by category.
- Reduce false positives: You’ll find that recurring shadows and other lighting effects are learned as a category. There are also activities that happen as one-offs, which may be classified as “Unrecognized.” You can turn off alerts and filter out motion markers for these categories so you don’t receive as many “false positive” notifications.
- Find the needle in the haystack: Look out for the color-coded motion markers on your timeline that occur infrequently — they could be interesting events you might want to save as a clip!
Let’s geek out for a minute…
There’s a lot of interesting technology behind Activity Recognition. For each activity, moving objects are segmented and a rough estimate of their paths is computed, among other statistics. Our system then uses clustering algorithms to find and group typically occurring variations of these statistics. Complicating matters, consider the scale of learning these activities — our servers currently process upwards of 8 million activities per day. To handle this mass of data the learned models for your camera are updated online, i.e., sequentially, with each new activity. An added bonus of this type of learning is that Activity Recognition gets progressively better over time.
Tell Us What You Think
As you might expect with any beta, Activity Recognition may not work perfectly all the time. Some events will not be recognized when it seems like they should be and some categories may be difficult to interpret.
We’re constantly improving Activity Recognition, so we’d love to hear what you think is useful, useless or confusing. Give it a try and send us your feedback.