The Port of Barcelona has been awarded one of the World Smart City 2021 awards, specifically in the Mobility section, for a solution aimed at mobility management at the cruise terminals. This smart mobility system has been developed with the collaboration of the companies Deep Solutions and Delonia (Port de Barcelona press release).
Deep Solutions successfully completed its mobile phone detection system with the collaboration of Generalitat’s Transit Service. The neural network has been trained with nearly 100.000 images of vehicles, and approximately 1% of then had people using their mobile phones while driving.
With the use of Deep Traffic assistant system, the traffic department can multiply by 50 its capacity to detect traffic violations without the need to increase personnel or working hours. The technology leap forward has been well received by the traffic authorities.
The visual review of random traffic images is a very monotonous task, and it often leads to the loss of the inspector’s attention. Deep Traffic orders the images so that the inspector sees the images with highest probability of containing an infringement first. This greatly increases the detection capabilities of the inspection personnel and maximizes the performance of his time.
It has been developed from the EfficientDet architecture, and trained with nearly 100,000 tagged images.
Deep Solutions successfully implemented the proof of concept of the access control system with thermographic monitoring for the 2020 Formula 1 Grand Prix at the Circuit de Catalunya.
In this pilot we participated together with CTTI, TSYSTEMS and Bioidenti to implement a facial and temperature recognition system, all in one, that allows to manage the control of accreditations and also to alert in case of high temperature.
Deep Solutions obtains the second prize in the Smartcatalonia Challenge of 2019 with the application of counting of vehicles and people that enter the port facilities.
Yolo version 5, this time for PyTorch, is now available: https://github.com/ultralytics/yolov5.
According to its publication, it shows better performance than EfficientDet.
We at Deep Solutions are using it in the development of the forklift detector and we are satisfied with the result. We have trained the network in PyTorch and then adapted it to TensorRT for use in an Nvidia DeepStream pipeline giving excellent results.
EfficientDet is born, a network oriented to the detection of objects that can be adapted to the needs of each type of platform and achieves impressive results.
The paper (Mingxing Tan, Ruoming Pang, Quoc V. Le, Google Research, Brain Team) can be found here.
Published interesting study where AI helps to classify IVF embryos as good, fair, or poor, based on the likelihood each would successfully implant.
Scientists of Cornell University have trained a Inception-V1 network with more than 12,000 images achieving an interesting AUC of 0.988 in their evaluation dataset.
Paper can be found here.
Interesting application of Generative Adversarial Networks (GAN), the same ones that allow us to create those popular ‘fakes’ where someone’s face is replaced by that of a famous person, or a video is modified so that a person seems to be saying something they did not say.
This article shows how the use of deep learning techniques to decode mental images can generate truly impressive results.