Ajackus Partnered with Phasorlab developed a custom model that predicts the depth of a 2d image up to 10 meters.
PhasorLab uses High-Precision Synchronization Technology, Hyper Sync Net. It is capable of maintaining time synchronization of better than 1 nanosecond and frequency synchronization of better than 1 ppb, which is key to achieving a centimeter-level target tracking accuracy for network-based positioning systems suitable for both indoor & outdoor applications.
PhasorLab wanted a computer vision solution that augments proprietary high-precision network-based positioning. They also wanted this solution to be usable with a single CCTV camera.
We divided requirements into two major modules,
For body detection, we had to detect a human in an image. We came across many helpful libraries that would help us do this. Mask RCNN is a good object and instance segmentation library. It gives us good object detection results and we were able to pick out the humans detected in the images.
For depth detection, we needed to determine the distance of the human from the camera. Thus depth detection needed to be done on humans from the image input. After a good amount of research on the depth detections algorithms available online, we decided to use DenseDepth algorithm model for depth detection.
Once the object detection and depth detection is done on the image, we extract only the depth of the humans from the image. In order to do this, we used the instance segmentation of the Mask RCNN object detections & extracted the depth detection of the human in the image.