YOLO object detection
This investigation is trialling AI object detection on 'edge' based AI devices.
The video shown is running on a Snapdragon 835 development board doing inference on a 720p video stream.
The software is a version of Darknet with the Yolo 9000 dataset.
Extensive modifications of the code are required to target the GPU/DSP unit of the Snapdragon device.
Further development will be targeting the Jetson TX2 developement board.
This should proceed more smoothly as the TensorFlow graphs are more easily converted to run on this device
due to the Jetson supporting CUDA.
Usage of the yolo 9000 data set is not optimal here. Notice the detection of flying cows, elephants and other artifacts.
See 2:35 for trees being recognised as elephants etc with a higher than expected probability.
Training a new model on with fewer classes is to be implemented.
