Our AI & Data Science Case Studies

Automatic human body measurement

Project challenge

The objective of this project was to develop a method for accurately measuring the dimensions of the human body. This precision is crucial for creating realistic and personalized 3D avatars, applicable in diverse fields.

Our approach

Integration of Advanced Pose Estimation

We utilized OpenPose, a cutting-edge pose estimation technology, to accurately capture and map the human body's posture and movements.
Segmentation of Body Parts

Employing BodyPix, we segmented the body into distinct parts. This segmentation was vital for isolating and precisely analyzing each section of the body.
Application of Curve Fitting Techniques

To model the body's dimensions accurately, we implemented sophisticated curve fitting methods. This step was critical in ensuring the measurements closely mirrored the individual's actual body shape.

Car presence detection and discrimination

Project challenge

Our primary challenge was to develop a validation module capable of reducing false positive violations generated by in-field sensors in parking violation detection systems.

Our approach

1 Designing the Validation Module

We created a specialized module focused on accurately identifying and validating parking violations, distinguishing between true and false positives.

2 Technology Integration

Multilabel Image Classification: This technique was used to analyze visible car surfaces, a key factor in identifying parking violations accurately.
Siamese CNN for Car Matching: We employed Siamese Convolutional Neural Networks to match cars effectively, ensuring precise vehicle identification throughout the monitoring process.
Motion Detection: Incorporating motion detection technology allowed us to track vehicle movements, an essential component in determining parking violations

Results

Enhanced Accuracy

The implementation of our solution led to a significant reduction in false positive violations, thereby increasing the system's accuracy.
Improved Reliability

The enhanced system reliability fostered greater trust and efficiency in parking violation enforcement.
Operational Efficiency

Our solution streamlined the process of parking violation detection, making it more efficient and effective for users and enforcement agencies.

Detection of cars on overhead images

Project challenge

The primary aim of this project was to calculate the density of cars in a chosen region. This analysis was crucial for understanding traffic patterns, aiding urban planning, and developing smart city initiatives.

Our approach

Implementation of Object Detection

We utilized YOLO (You Only Look Once), an advanced object detection system, to accurately identify and count cars within various urban images.
Processing Map Tiles

Using the Google API, we processed map tiles to obtain detailed geographical data of the chosen region. This step was essential for accurately pinpointing the locations where car density needed to be calculated.
Map Visualization

To effectively present our findings, we integrated Leaflet.js for map visualization. This allowed us to display the car density data on interactive maps, making the information easily accessible and understandable.

Try Our Demo

Fabric Defects Detection

Project challenge

The goal of this project was to create a reliable, automated tool capable of identifying various defects in fabrics used in polymer composite material production. This tool aimed to enhance quality control, reduce waste, and improve the efficiency of the manufacturing process.

Our approach

Dataset Preparation

We generated a synthetic dataset using advanced image processing techniques. This dataset was tailored to represent a wide range of fabric defects, providing a robust foundation for training our Machine learning model.
Model Development and Training

We employed the U-net architecture, a powerful image semantic segmentation tool, within the TensorFlow framework. This approach allowed for precise segmentation and identification of defects in the fabric materials

Football player performance evaluation

Project challenge

The primary goal of this project was to develop a sophisticated system capable of quantifying athletic performance with high accuracy. This involved tracking and analyzing the movements of football players during games to provide valuable insights into their performance.

Our approach

CNN-Based Field Marking Recognition

We utilized Convolutional Neural Networks (CNN) to identify key points and lines on the football field. This technology allowed us to establish a reference framework for accurately tracking the players' movements.
Custom Camera Calibration

Fully automatic camera calibration algorithm for the entire video that uses keypoint and line data for all frames of the video. This allows accurate camera calibration even on frames with a high zoom, where few keypoints and few lines are visible.
Object Detection and Tracking

Our system employed advanced object detection and tracking algorithms. These were instrumental in monitoring the athletes' positions and movements throughout the game.
Pose Estimation

We implemented a lightweight pose estimation model. This model was designed to recognize the projection point of an athlete's pelvis onto the field, a key metric in assessing movement efficiency and agility.

Results

Accurate Performance Metrics

The system reliably calculated essential metrics like speed and acceleration, offering coaches and players valuable insights for training and strategy development.
Efficient Resource Usage

Despite its complexity, the pose estimation model was optimized to require less than 200 MB of GPU memory and function effectively at low resolutions, ensuring its applicability in various environments without the need for high-end hardware.
Enhanced Training and Strategy

The insights provided by our system have been instrumental in enhancing training regimes and in-game strategies, directly contributing to improved athletic performance.

Online exercises quality evaluation

Project challenge

The main aim of this project was to develop a web application for home workouts that evaluates exercises in real-time. The application should allow users to select an example exercise and then attempt to replicate it, providing immediate feedback on their performance.

Online exercises quality evaluation: Project approach

Our approach

Integration of Pose Estimation

We implemented pose estimation using MediaPipe POSE to accurately track and analyze the user's movements during exercises.
Applying Time Series Comparison

To assess the accuracy of the user's exercise performance, we utilized the Dynamic Time Warping (DTW) algorithm.
Enabling Real-Time Video Stream Processing

The application was developed to process video streams in real time, ensuring that users receive immediate and relevant feedback as they perform the exercises.

Palm recognition

Project challenge

The project aimed to develop a module capable of validating user-submitted hand photos for an online palmistry service. The module's key objectives were to ensure that the hand is present in the image, the entire hand is visible, it is correctly oriented, and the lines on the hand are clearly discernible.

Our approach

Photo Validation Criteria Establishment

We defined specific criteria for photo validation: presence of the hand, visibility of the entire hand, correct orientation, and clarity of hand lines.
Technology Utilization

We employed TensorFlow to develop a heatmap-based points and lines recognition system. This technology was chosen for its precision in identifying and analyzing specific features in images, which is crucial for validating the key aspects of hand photos.

Realtime playing card recognition

Project challenge

The primary objective of this project was to develop a playing card recognition module. This module is a key component of a system designed for real-time analysis of casino games. Its purpose is to accurately identify and classify playing cards during live games, enhancing the monitoring and analysis capabilities within a casino environment.

Our approach

Creation of a Synthetic Dataset

We developed a synthetic dataset specifically tailored for training deep learning models. This dataset was instrumental in training our object detection system to recognize a wide variety of playing card designs.
Implementation of YOLOv4 Object Detector

We utilized the YOLOv4 object detection framework, known for its efficiency and accuracy in real-time object recognition tasks.
Real-Time Video Stream Processing

We integrated our system with real-time video stream processing capabilities. This allowed the module to analyze and identify playing cards in live scenarios, a critical requirement for real-time game analysis in casinos.

Results

High Recognition Accuracy

It can recognize cards with almost any design, demonstrating a high level of accuracy and adaptability.
Versatile Detection Capabilities

The module is capable of detecting playing cards from any angle and in various environments, showcasing its effectiveness in diverse and dynamic casino settings.
Performance Under Low Image Quality

Even with low image quality, the system can successfully detect and classify playing cards, ensuring reliability and consistency in challenging visual conditions.

Recognition of the area of asphalt pavement from satellite images

Project challenge

The objective of this project was to develop a sophisticated system for calculating the area of asphalt pavement required for parking lots. This calculation is crucial for accurately estimating the cost of asphalting work.

Our approach

Extraction of Parking Boundaries

Utilizing the Overpass API, we extracted parking boundaries from OpenStreetMap data, ensuring a reliable foundation for our calculations.
Advanced Image Segmentation

We implemented semantic image segmentation using a u-net convolutional neural network (CNN) built with TensorFlow. This allowed us to accurately distinguish between asphalted and non-asphalted areas within parking lots.
Map Tiles Processing

Leveraging the Google API, we processed map tiles to obtain detailed geographical data, essential for accurate area calculation.
Interactive Map Visualization

To enhance user experience and accuracy, we integrated Leaflet.js for dynamic map visualization, enabling users to interactively view and modify parking lot data.

Results

As a result, we developed a microservice and a web application that enabled users to:

Navigate maps and automatically detect parking contours.
Search for parking lots by address.
Obtain the area of asphalt pavement for selected parking lots, with AI recognition of unpaved areas like grass patches and buildings within the parking lot's outer contour.
Edit the results of unpaved area recognition.

Demand prediction based on historical data

Project challenge

The goal of this project was to create a predictive model that could accurately forecast the demand for transportation services. The model aimed to consider various factors such as seasonality, national holidays, and weather conditions, enabling transport companies to plan the rational placement of vehicles and improve service efficiency.

Our approach

In tackling the challenge of accurately predicting transportation service demand, our approach was multi-faceted and data-driven. We focused on three key areas:

1 Data Collection and Analysis

Historical data on transportation service demand was meticulously analyzed, focusing on patterns influenced by seasonality, national holidays, and weather conditions.

2 Model Development

An autoregression model was developed to capture the time-dependent nature of transportation demand.
Random Forest regression, known for its effectiveness in handling complex, non-linear data patterns, was employed to enhance the model's accuracy.

3 Integration of Weather Data

A third-party API was utilized to integrate real-time weather data into the model, acknowledging the significant impact of weather conditions on transportation demand.

Evaluation of human body hydration

Project challenge

The goal was to create a model using data from a biomedical wearable device to accurately predict the hydration level of individuals. This tool aimed to assist patients with kidney failure or ESRD in monitoring their condition and planning their dialysis sessions more effectively.

Our approach

In developing a model for assessing human body hydration levels, our approach was comprehensive, encompassing various stages. Here's a breakdown of each key stage:

Data Collection

Data was collected from patients wearing the device during dialysis sessions and overnight, including device signals and ultrafiltration volume from the dialysis machine.
Data Preprocessing

The preprocessing steps included data synchronization, temperature-based data selection, low-pass filtering, outlier detection using an Isolation Forest model, and signal smoothing.
Feature Engineering

Based on skin reflection data at different wavelengths, 92 initial features were selected for analysis, informed by literature reviews and laboratory experiments.
Model Development

Two regression models, Theil-Sen and multi-layer perceptron (MLP), were considered. Models were trained individually for each patient, as well as in general and gender-specific formats.

Results

We meticulously analyzed the outcomes to ensure the model met the high standards required for medical applications, especially for patients with critical needs. Here's a detailed overview of the key results:

Model Performance

The MLP model outperformed the Theil-Sen regression, exhibiting a lower root mean squared error. Individual patient models yielded more accurate results compared to generalized models.
Accuracy Metrics

The average R2-score across patients was 0.8, with accuracy metrics ranging from 0.1 to 0.15, indicating high precision as evaluated by the clinical team.
Real-World Applicability

The model's predictions on overnight data, though not directly testable for accuracy, were deemed reasonable and consistent with clinical expectations.

Evaluation of human body
hydration: Results

Realtime muscle fatigue estimation

Project challenge

The project aimed to create a system that could instantly analyze EMG data to determine the degree of muscle fatigue during exercises. This tool was intended to aid research in sports medicine, using the Bitalino MuscleBIT EMG device for data collection.

Our approach

In developing a real-time muscle fatigue estimation tool, our approach encompassed the following steps:

1 Data Collection and Preprocessing:

EMG data was recorded using the Bitalino MuscleBIT device and Open Signal application, capturing raw signals at 1000 hertz during maximum repetition exercises.
Preprocessing included mean normalization and bandpass filtering, though the model proved insensitive to noise, rendering frequency filtering unnecessary.

2 Muscle Fatigue Detection Methodology:

A convolutional neural network (CNN) was employed to process EMG data slices, focusing on slices with significant EMG amplitude.
Ground truth for muscle fatigue was established based on the assumption that initial exercise repetitions indicated zero fatigue and final ones indicated maximum fatigue.

3 Model Development and Alternative Approach

The CNN model consisted of layers designed for effective pattern recognition in EMG data.
An alternative approach using EMG spectral analysis was also explored, assessing muscle fatigue through shifts in signal spectrum frequency.

Results

As a result, the binary accuracy of EMG slice classification for the validation data is 91 %, and for test data is 89 %. Compared to the traditional approach based on shift of the signal spectrum, the CNN-based approach gives smoother dependences on the degree of muscle fatigue over time and is better consistent with the subjects' feelings of muscle fatigue

Structural health monitoring of carbon fiber aircraft structures

Project challenge

The goal of this project was to create a machine learning-based system that could process sensor data obtained from electrical sensors embedded in carbon-fiber airplane panels to detect low-velocity impact damage. The system was designed not only to recognize patterns indicating damage but also to estimate the damage's location and size. This innovation aimed to reduce maintenance costs and increase the safety of aircraft operations.

Our approach

To achieve our goal of developing a reliable damage detection system for aircraft, we focused on the following critical stages.

System Development

A machine learning system was developed using TensorFlow. This system was designed to analyze sensor data and identify patterns corresponding to structural damage.
Damage Estimation

The system was equipped with capabilities to accurately estimate the location and size of detected damage, providing essential information for effective maintenance.
Synthetic Dataset Creation

A synthetic dataset, essential for training the machine learning model, was generated using Finite Element Method (FEM) simulation. This dataset ensured comprehensive coverage of potential damage scenarios.
Full-Scale Testing

The system underwent full-scale testing to validate its effectiveness in a real-world setting.

Custom camera calibration algorithm

Project challenge

The primary goal of this project was to create an advanced algorithm as part of a football analytics system to accurately determine the positions of athletes on the field. This tool was designed to consider key field markings, such as yard lines, to enhance the precision of player positioning data.

Our approach

Our approach combined advanced computational techniques with a deep understanding of American football dynamics. Here's a closer look at the key components of our approach and implementation.

Algorithm Development

We developed an algorithm that leverages field keypoints and yard lines to determine player positions with high accuracy.
Field Marking Recognition

We implemented a system to recognize and interpret field markings, including points and lines, crucial for accurate player positioning.

Technologies used

Perspective projection with Pinhole camera model, using OpenCV
Solution of the nonlinear optimization problem to determine parameters of a camera based on the result of field marking recognition (points and lines) using SciPy library
Custom Object tracking algorithm for yard lines

Estimation of the machining time on a CNC machine

Project challenge

The goal of this project was to decrease the time needed to estimate the production cost of each detail with CNC machines. Previously it took several days to do this. We decided to create an automated system that could instantly provide customers with estimates for machining mechanical parts on CNC machines. This system aimed to determine the type of equipment required, the size of the blank, the list of operations for machining the part, the number of tool changes, and the machining time for each operation.

Our approach

We focused on developing a comprehensive solution that integrates advanced technological capabilities with user-friendly access. Here's a detailed look at each aspect of our approach and implementation:

Our approach

We focused on developing a comprehensive solution that integrates advanced technological capabilities with user-friendly access. Here's a detailed look at each aspect of our approach and implementation:

System Development

Developed an automated estimation system integrated into a web platform, allowing customers to upload part designs and receive immediate cost and time estimates.
Feature Analysis

Implemented shape recognition and classification algorithms to analyze the 3D designs of mechanical parts and identify key machining features.
Operational Calculations

Created a methodology to calculate the necessary operations, including the number of tool changes and the time required for each operation, based on the part's design.

Estimation of the machining time on a CNC machine: Our approach

Estimation of the machining time on a CNC machine: Our approach 2

Logistics system for technical service

Project challenge

The project aimed to develop a sophisticated route optimization system for multiple vehicles. This system will consider specific time windows, adhere to constraints on the length of the working day, and determine the most efficient dates for scheduling work orders.

Our approach

In this project, we adopted a multifaceted approach, combining advanced algorithmic strategies with high-performance computing and data analytics to tackle the complex challenge of multi-vehicle route optimization. Here's a closer look at the key components of our approach and implementation:

Algorithm Development

A route optimization solution was created using a combination of the Simulated Annealing Algorithm and the Greedy Algorithm, implemented in C++ for computational efficiency.

Computational Strategy

The project leveraged multiprocessor computations to handle the intensive computational demands of route optimization.

Time Matrix Calculation

The Open Source Routing Machine (OSRM) was employed for the initial calculation of the time matrix, essential for route planning.

Traffic Data Integration

A regression model trained on historical traffic data was used to adjust the time matrix, accounting for traffic variations and enhancing route accuracy.

Our AI & Data Science Success Stories

Artificial Intelligence

Data Science

Automatic human body measurement

Detection of cars on overhead images

Online exercises quality evaluation

Football player performance evaluation

Car presence detection and discrimination

Realtime playing card recognition

Fabric Defects Detection

Palm recognition

Demand prediction based on historical data

Structural health monitoring of carbon fiber aircraft structures

Realtime muscle fatigue estimation

Evaluation of human body hydration

Logistics system for technical service

Custom camera calibration algorithm

Estimation of the machining time on a CNC machine

Project challenge

Our approach

Integration of Advanced Pose Estimation

Segmentation of Body Parts

Application of Curve Fitting Techniques

Project challenge

Our approach

1 Designing the Validation Module

2 Technology Integration

Results

Enhanced Accuracy

Improved Reliability

Operational Efficiency

Project challenge

Our approach

Implementation of Object Detection

Processing Map Tiles

Map Visualization

Project challenge

Our approach

Dataset Preparation

Model Development and Training

Project challenge

Our approach

CNN-Based Field Marking Recognition

Custom Camera Calibration

Object Detection and Tracking

Pose Estimation

Results

Accurate Performance Metrics

Efficient Resource Usage

Enhanced Training and Strategy

Project challenge

Our approach

Integration of Pose Estimation

Applying Time Series Comparison

Enabling Real-Time Video Stream Processing

Project challenge

Our approach

Photo Validation Criteria Establishment

Technology Utilization

Project challenge

Our approach

Creation of a Synthetic Dataset

Implementation of YOLOv4 Object Detector

Real-Time Video Stream Processing

Results

High Recognition Accuracy

Versatile Detection Capabilities

Performance Under Low Image Quality

Project challenge

Our approach

Extraction of Parking Boundaries

Advanced Image Segmentation

Map Tiles Processing

Interactive Map Visualization

Results

Project challenge

Our approach

1 Data Collection and Analysis

2 Model Development

3 Integration of Weather Data