Traditionally, the performance of AI models for “real-time” object detection is directly linked to the computing power and storage capacity of the hardware used to run the model. This factor heavily limits the application possibilities of Neural Networks on the edge of the IoT cloud.
At the same time, embedding a Neural Network into an IoT system has a series of advantages. By eliminating the need to send data to a remote computer for classification, embedded systems enable applications which are truly “real-time” and can function without the availability of mobile networks. This creates countless opportunities for safety critical and GDPR compliant applications.
I was part of a project that explored the possibilities of optimizing current lightweight models for traffic detection. In cooperation with researchers at the Karlsruhe Institute of Technology, our team at NovelSense tested and compared different optimization procedures to evaluate the possibility of creating Neural Nets with a superior performance and less than average size. The main case was the classification and counting of E-Scooters in an Urban traffic setting. In the following video Yexu Zhou (TECO, KIT) explains the scope of the project and current results.
In the project we generated training and test data utilizing the possibilities of a Generative Adversarial Network (GAN). Subsequently the main goal was to test different modes of reducing the model size without comprimizing the recognition performance. Intermediate results show that we were able to achieve a better than average detection accuracy with less than half the model size of state of the art (lightweight) classification models. This was achieved by optimizing the architecture of the model and pre-processing procedure.
The results show that there is still a big potential in optimizing recognition Software without the need for high performance hardware architectures. The project will continue to explore these potentials and aims to create innovation in the area of embedded Artificial Intelligence.