Artificial Intelligence in Road Traffic Measurement.

Neural networks processing images and object tracking algorithms.

How neural networks recognises objects and classify images?

Image classification by artificial intelligence is the first step in creating an AI model that can recognize multiple objects simultaneously. Image classification, in simplified terms, involves assigning a name to the most prominent object in the image, essentially "labeling" the image. For example, an image of a boat might be labeled as "boat 99%."

The beginnings of image classification involved the use of fully connected neural networks (fully connected networks trained on a dataset of handwritten characters called the MNIST dataset). A breakthrough discovery was the design of Convolutional Neural Network (CNN) layers, which combined trainable networks with image filtering. It was only with the use of networks based on CNN layers that the first high-quality classifiers were achieved, such as the AlexNet network trained on the ImageNet dataset.

Fragment of the MNIST training dataset containing images of handwritten digits.

The architecture of the AlexNet network.

Each image is classified with labels and a % likelihood value.

An attempt to recognize a dog's breed from a photo using artificial intelligence.

YOLO is a groundbreaking architecture

The breakthrough architecture was the YOLO v1 network designed by Joseph Redmon. It not only allowed for image classification but also for the simultaneous detection of multiple objects. YOLO processed images quickly (for its time) and achieved 30 frames per second (measured on an NVIDIA Pascal Titan X card).

YOLO in subsequent versions, v2, v3, and v4 created a sensation in the open-source and business communities. The authors made YOLO fully available under the MIT license, which allows for its commercial use, development, and modifications.

Training AI on 40,000 Images

At AISP, we use YOLO for recognizing 22 classes of objects. We've collected a dataset of over 40,000 images, which allows us to properly train the neural network. Each image has been manually annotated with rectangles and labels for the objects present in them. We used yaya software for this annotation.

Our images include adverse weather conditions, fog, occlusion encounters, noise, and other challenging circumstances. All of this is to ensure that we can recognize objects in any situation!

Identity recognition

Recognizing the class of an object is the first step. How can we confirm the identity of an object across multiple video frames? How do we know that it's the same car that entered the intersection from the left lane 2 seconds ago? This requires object tracking methods.

Multi-object tracking a continuously evolving field of study, sees new algorithms emerging year after year. Many well-known algorithms like SORT and DeepSORT achieve good results, but counting with 97% accuracy requires something more. That something is our proprietary algorithm called VRT (Vehicles Real-Time Tracking).

Relations classification and cross-sections

The movement of a tracked vehicle is a trajectory of positions (x, y, z) over time t. How can we recognize in which relation the vehicle was moving or which cross-section of the road it crossed?

Algorithms of linear algebra and special metrics are used for this purpose. For example, the Frechet distance allows us to calculate the similarity between curves. This way, we can clearly determine which movement curve is most similar to a given relation.

With AITracker at AISP, we can use these methods to detect intersections of a vehicle's trajectory with a cross-section, entrances/exits to/from restricted areas, or assign specific movement curves to relations at intersections.

AISP - AI and image recognition specialists

AISP consists of specialists in image recognition, artificial intelligence, and object counting from video recordings. We are eager to hear how you would like to use our system and how we can meet these challenges and offer the best solution for you.

Let's discuss a solution for you