The World of Smart Mobility
Today we have 33 megacities in the World, 55% of the global population is urbanized, and the predictions are that by 2030 this urbanization will reach 60%. The world’s population will swell to 10 billion by 2050. Cities are getting overcrowded, and in many cases, it’s a nightmare to move inside them.
People going to/from work by trains, trams, cars, taxis, bikes, electric bikes, electric boards, subway, trucks delivering goods to various spots, increasing number of drones and autonomous vehicles, the list is long.
That’s why in the near future the mobility sector needs to disrupt, and probably soon it will look totally different than we know it today. Cutting-edge technologies in public transport are sth we need to optimize the way we move and to avoid transportation paralysis.
In order to keep up with upcoming changes and technologies in the mobility sector, Tooploox team is going to visit the MOVE2020 conference in London. We want to see what are the newest trends, and where the World of smart mobility is heading.
See you all there!
And how do we help to shape the future of mobility?
Light – the camera that helps cars see as humans
Improving accuracy of stereo depth
Apart from creating more vivid and more detailed HDR images, using several cameras in smartphones or compact cameras allows to exploit stereovision and calculate the depth of the image. This depth can be used for many different purposes, among others nice effects applied to different parts of the image based on depth segmentation.
The main goals tackled by Tooploox are improving the sensor relative geometric calibration and improving the quality of stereo depth reconstruction. The first issue is related to the fact that the individual cameras are not glued to any surface and are prone to micro movements when the device is used. It turns out that these movements can be significant and may require recalibration of the whole layout for each acquired image.
The project is planned for an indefinite length, with the first set of deliverables due to be completed after around 3 months. The work so far has been focused on the issue of online calibration: what are the theoretical and practical gains from using it. The next focus item will be improving the quality of stereo depth reconstruction.
We have successfully implemented the following modules:
- A framework for camera calibration testing w.r.t. synthetic 3D points
- A framework for online camera calibration configuration comparison using both synthetic and real-world data
- Tools for 3D point cloud storage and visualization
- A parameterized online calibration module using Ceres nonlinear optimization, allowing to parameterize the bundle adjustment problem for camera intrinsics, camera extrinsics, point position and point depth
- Conclusive proof that online calibration is useful for improving stereo depth quality and a quantitative evaluation of this gain
- Conclusive proof that correctly set up online calibration is capable of improving camera parameters that have been perturbed
- Improving the average reprojection error from 8 to 0.9 pixels for synthetic images
- Improving the average reprojection error from 1.5 to 0.9 pixels for real-world images
Voyage – autonomous driving vehicle that reads passengers emotions
We developed an in-car system for autonomous cars that counts passengers and analyzes demographics. Our solution receives information from a single camera placed inside the vehicle through ROS (Robot Operating System) and handles it through a deep CNN module. It successfully deals with occlusion, passenger movement inside the car as well as very difficult lighting conditions.
Expected goal of our project was to create a system with the following modules: face detection/recognition, age and gender classification, face expression recognition.
- Backend integration with ROS
- Integration with RGBD camera
Project duration: 6 months
Competences and technology: ML/CV expertise
Face detection module builds on a multi-task cascaded convolutional network (MTCNN) that consists of candidate window proposal network and a more complex network that rejects non-face windows, refines their position and performs non-maximum suppression to merge highly overlapped candidates. Training is performed in a three-stage manner that allows for iterative improvement of face detection bounding box.
Age regression problem is posed as a deep classification problem followed by a softmax expected value refinement with WideResNet architecture as deep feature extractor.
Facial expression recognition
We have developed a robust face alignment and emotion recognition method based on a deep neural network architecture called EmotionalDAN which outperforms state-of-the-art by 5%. As a result of this work we have academic record at top tier conferences for Computer Vision (https://arxiv.org/abs/1805.00326 , CVPRW 2018)
AI powered helicopter patrols for Vegetation Management and Pole Top inspections
An object detection task where we want to have an automatic system that ingests a picture of an electric pole top taken by a drone and detects and annotates one of 16 types of objects that the pole can consist of.
Produce a Proof-of-Concept web service for analyzing images. The goal of the service is to be used as a part of a larger system in order to show future capabilities to potential investors and clients.
A Python-based ReST web service was deployed on the client’s Google Cloud Platform. The service ingested an image and produced a JSON output with bounding boxes of detected objects along with their names. The service used the Mask R-CNN algorithm trained on images provided by the client.