Computer vision, our uses and our challenges
Computer vision refers to an artificial intelligence technique that allows analyzing images and videos captured by a camera. Artificial intelligence is capable of analyzing an image, understanding it, and processing the resulting information. The current performance of sensors, relative to video quality, gives artificial intelligence algorithms the ability to see the world. By categorizing objects, their movements in space, and analyzing visual factors corresponding to a specific situation, AI will be able to analyze in real time and send information for humans to access and potentially make a decision.
For example, our algorithm has been trained to detect and categorize fires in order to alert the competent authorities in real time for a faster and more efficient intervention.
Communities and infrastructures are already equipped with cameras for a multitude of uses: counting mobilities, studying flows, protecting buildings and public installations, observing traffic rule violations, protecting the immediate surroundings of businesses in areas particularly exposed to risks of assault or theft... Without the overlay of artificial intelligence, it is up to the human eye to see and analyze all situations. However, every year the amount of information accumulated by humanity doubles and only 70% of this data is perceived by human vision. When we reach the limits of the cognitive capacities necessary to process all this information, our brain is no longer capable. Due to its evolution, it operates a biased analysis of reality in order to optimize its energy expenditure. Video analysis in artificial intelligence will thus relieve and support the human eye.
One of the main challenges of computer vision is related to the complexity of the visual world. An object can be perceived from multiple angles, in various lighting conditions, partially obscured by other objects... A true computer vision system must be able to perceive content in any of these situations and extract information from it. Hence, computer vision represents a real challenge.
Introducing new technology and object perceptions raises new legal, ethical, and political issues. As cameras are deployed in public and private spaces, questions about the individual freedoms of people arise. As soon as there is a human presence or an object enabling an indirect association with a person, image (photos or videos) capture and processing are governed by the GDPR. At XXII, we do not analyze personal data: we distinguish silhouettes that we associate with an object category (human, dog, car, bicycle...) without ever using biometric data. Beyond object categorization, some of our data processing aims at analysis and optimization (assembly lines, mobility flow analysis...). In our algorithmic analyses, numerous data, including personal data, can be considered as being in a "dormant" state. They are captured by the camera but do not enter the artificial intelligence analysis scheme.
Data protection impact assessment.
Why?
Having a dataset of photos or videos is essential for AI. Without it, the technology cannot exist. Vision requires a large amount of image and video data to learn, test, and analyze models. This is what is called deep learning: deep learning consists of algorithms capable of mimicking the actions of the human brain through artificial neural networks. The networks are composed of tens or even hundreds of "layers" of neurons, each receiving and interpreting information from the previous layer. As soon as we build a dataset that allows the machine to recognize and classify objects, it is done in compliance with the GDPR and in some cases, may be framed by a data protection impact assessment.
Creation of a Database
We create specific databases for exclusive internal scientific research purposes at XXII, in compliance with the relevant people and legislation. In order to minimize impacts on people related to the processing of potential personal data, we are currently working on setting up dataset generation using synthetic data. The constituted datasets are targeted for a specific use case (for example, learning a new object) and varied (for example, challenging the detection of an object in various situations). Also, our datasets are compiled in strict compliance with the GDPR: using open-source solutions, with the consent of the individuals concerned, in partnership with various actors, and by pseudonymizing the data.
Annotation
In the development of an artificial intelligence algorithm, certain processes such as learning require a corpus of annotated data. This is why, internally, the first phase of research and development focuses on annotating the collected data (object classes, camera angles, weather conditions...).
Training
The goal of computer vision is to allow the computer to see, analyze, and understand one or more images received by the system. This is why we teach our artificial intelligence models, often deep neural networks, on a corpus (thus one or more databases) containing annotated examples. Whether for training, or retraining (specializing, correcting a bias, etc...) our algorithms, visual data (images, videos) is necessary.
Evaluation
The evaluation, or testing, of our models aims to measure our performance according to a set of predefined criteria. The evaluation can concern the entire model or part of the model. We are particularly interested in the true positive rate, false positive rate, true negative rate, and false negative rate. Based on these measurements, we can build more readable indicators (sensitivity, specificity, etc).
These measurements are made by object classes and combined to provide an average score. A class corresponds to a more or less precise object typology (human, bicycle, cyclist, car, vehicle, ...). They are important because they allow us to understand the limitations of our models and communicate them to our users. These tests can be performed on pseudonymized data, using proprietary software or developed internally.
One of the major challenges of artificial intelligence is creating algorithms free from any bias. We speak of bias when AI favors one situation over another. For example, is there a difference in detection between blue and red cars? To counteract potential biases, numerous datasets are needed to enrich the models so that they are varied and diversified. Our R&D teams have implemented a continuous process of developing, learning, and testing our algorithms to address potential biases.
Furthermore, the diversity of databases is also crucial for evaluating our performance according to the angles from which images are viewed. The diversity of angles included in our databases will allow us to determine if, according to certain camera angles, our algorithms detect the sought-after object more or less effectively.
GDPR & ethics at XXII
CORE, a decision support tool
Computer vision processes video streams containing personal data, therefore, it is essential for us to be respectful of the GDPR and complementary texts. We also reaffirm our position as a decision support tool. Under no circumstances will we replace the human eye; our solution complements it. Indeed, our platform is a decision support tool for operators. It does not trigger, for example, any automated procedure following a committed or presumed infringement, and simply facilitates access to information already present within a security center. Therefore, the use of our solution is always framed by human intervention. Our platform respects the values upheld by the CNIL and observed throughout its position:
The necessity of our product lies in the utility of the device in view of the clearly identified objective.
The proportionality lies in the pre-existing presence of video surveillance. Our product does not impact the persons concerned in any way other than that intended by video protection or surveillance.
Data minimization occurs notably through the absence of personal data storage by XXII. Indeed, our software does not store any video stream to date and operates in real time.
Non-identification, our algorithms consider a person as a silhouette, thereby freeing themselves from any personal data, so it is the position of this silhouette in space that is analyzed. Finally, the analysis is carried out on a group of people and not arbitrarily. There is therefore no targeted analysis of an individual.