AI and Data Engineering Services to EBYTE

AI and Data Engineering Services

EBYTE

Jun 2016 – Jun 2019

At AIWORX Company, we are proud to have provided AI and Data Engineering services to EBYTE, a remote company based in New Zealand. Our collaboration encompassed a range of innovative projects, including the development of a Street Planning System for Neurontix USA. This system features an analytics dashboard for traffic planning over CCTV infrastructure with wide-angle cameras on a three-lane street. It meticulously tracks every vehicle passing through each lane, identifying their type, parking spot density, and stop duration in long-term and delivery lanes. Additionally, the system recognizes delivery vans and brand advertisements on trucks. All traffic statistics are recorded into a database, with traffic analytics accessible via a ReactJS application that includes timestamped images and videos.

Our contributions to the Street Planning System were significant. We utilized transfer learning to improve publicly available vehicle detection models using our custom dataset, enhancing vehicle detection performance under low lighting and hazy conditions from 72% to 88%. One of the challenges we tackled was calibrating the cameras to preserve vehicle appearances across the entire image. Standard libraries often resulted in maximum distortion at the edges of the scene, so we developed a higher-order non-linear curve fitting for fisheye correction. This innovation increased overall vehicle detection under severely distorted conditions to 95%. Our technology stack for this project included Python, ReactJS, and Postgres Database.

In addition to the Street Planning System, we designed a produce counting system for HelloFresh New Zealand. This system accurately detects and counts up to 13 classes of vegetables and fruits on a conveyor running at normal speed, achieving an impressive 99% accuracy rate. The imaging system employs RealSense cameras, LED lights, and industrial-grade lasers, while the Nvidia AGX board manages image analysis with DeepSort as the real-time object detector and tracker. We also applied 3D reconstruction techniques to enhance accuracy and suppress false positives. The system’s outputs are displayed on an HMI, which can also provide various inputs to the control system and image analysis script. Our technology stack for this project included Python, DeepSort, OpenCV, RealSense, laser triangulation, LED control, AGX Jetson, and HMI. These projects showcase our commitment to delivering high-quality, innovative solutions that drive efficiency and accuracy in various industries.