Embedded Files
Deep learning powered by synthetic data
What exactly is people detection
People detection is the ability to find one or more humans in a video or image using computer vision. And while this technology has been around for a long time it historically suffered from low accuracy and false alarms. In the recent years deep learning (or AI) has helped improve on those issues considerably. The output of an AI model for person detection is a rectangle (or bounding box) that tells you were a person is located in the image. It also includes the confidence of the detection as well as a label showing it's a person.
The value of detecting people
People detectors reduce the false alarms found in older computer vision systems. This is because everyone used to rely on motion detection which can be very noisy. Any movement in a video, including an animal or tree in the wind, would be considered a person.
People detectors also form building blocks that are part of a bigger computer vision system. The information they produce is normally fed into a tracking algorithm, which allows a computer to follow a person over time. This is important because you cannot tell if someone is coming or going based on a single picture.
Applications using people detection
People detectors can be found in almost every smart system out there. Basically any system that interacts with us, whether it's for safety, security or comfort will try to detect people. But just to give you a few examples of the markets and their use-cases:
Smart cities: Pedestrian crossings, automated security gates,
Industrial safety: Forklift warning systems, wearing of personal protection equipment, robotic arms
Smart homes: Automatic AC and light control, doorbell notifications, home security
Perimeter security: Access control, intrusion alerts, people counting
What makes a good people detector
A people detector has many qualities that determine how well it's going to work for your application. Some of those qualities are based on what type of data the model was trained on. Others are determined by the type of hardware or camera it will run on. Picking the best combinations is often a game of trial and error.
Camera viewing angle
Most people detector models in the market only work for low elevation angles. So if you happen to place a camera too high up (top of a wall, building or from a drone) the models fail.
Because CVEDIA AI models are trained using synthetic data they work from any viewing angle, even directly top down. This gives you much greater flexibility when installing your cameras.
Location and environment
AI models are very sensitive to the data they train on, and people detectors are no different. If a model was trained only on indoor videos, it won't work well outdoors. The same goes for training on daytime images. Don't expect it to work well at night. Make sure you ask any vendor on what data they trained.
Our models are insensitive to the location, weather and light conditions they operate in. This is possible because we render our training data instead of collecting it. Giving you more robust models and a better user experience.
Accuracy vs speed
Depending on your AI chipset you'll want either small or big AI models. Small models are light-weight and can run on low-power device, but they take a hit in accuracy. Larger models run much slower and are suited better for powerful GPU's, but they provide higher accuracy.
Because this trade-off also depends on your application we provide a wide range of performance settings. Allowing you to pick the best model for your use-case.
Object size
The smaller the person you're trying to find, the more difficult it becomes for a people detector. But at the same time it allows you to use lower resolution cameras or cheaper lenses. It's another balancing game for which there is no one-size-fits-all.
To help you pick the best settings and models we have created a free software called CVEDIA-RT. It lets you to play with different settings to find out the optimal strategy for your application.
Report abuse