Recognizing motion does not end with naming the object. The system must detect it in the image, maintain its identity between frames, and interpret its trajectory.
How AITracker recognizes and tracks objects
From detection on a single frame, through vehicle tracking, to assigning a relationship or cross-section.
Briefly about the development of image recognition
A fragment of the MNIST training set, which contains images of handwritten digits.
Construction of the AlexNet network.
Each image is given a label and a probability value.
An attempt to recognize the breed of a dog from a photo using artificial intelligence.
YOLO is a breakthrough architecture
The network has become a breakthrough architecture YOLO v1 designed by Joseph Redmon. It allowed not only to classify photos, but also to locate many objects at once. YOLO processed photos quickly (for the time of its creation). It achieved 30 frames per second (measured on an NVIDIA Pascal Titan X card).
The subsequent versions of YOLO – v2, v3, and v4 – quickly spread to the open source community and commercial applications. The code is available under the MIT license, which allows development and business use.
AI training on 40,000 photos
At AISP, we use YOLO to recognize 22 object classes. We have collected a set of over 40,000 photos that allow us to properly train the neural network. Each photo was manually described with rectangles and labels of the objects in it. The description was used for this yaya software.
The collection includes fog, obstructions, noise and other difficult conditions. This allows the model to better cope with real road recordings.
Identity recognition
Recognizing the class of an object is the first step. How to confirm the identity of an object in multiple frames of a video? How to know that this is the same car that entered the intersection 2 seconds ago from the left entrance?
Need for an object tracking method.
Multiple Object Tracking (MOT) is a constantly evolving field. Algorithms such as SORT and DeepSORT provide a reference point, and AITracker also uses a proprietary VRT tracking algorithm adapted to traffic measurements.
Classification of relations and cross-sections
The movement of the tracked vehicle creates a position trajectory over time. The next step is to determine which route the vehicle was traveling in and what cross-section of the road it crossed.
This is achieved using linear algebra methods and metrics comparing curves, including the Fréchet distance. They allow you to determine which defined relationship the vehicle's trajectory best fits.
AITracker detects intersections of the trajectory with the cross-section, entries and exits from the area, and assigns the vehicle's movement to the relationship at the intersection.
Need help with your own material?
Download AITracker or go to the technical documentation to prepare your recording and measurement configuration.