This week had a lot of study regarding different ANN's and their performance on the RockPro. Upon doing some research, two NN's stood out as potential candidates from the list: YOLO Networks (You Only Look Once), and SSDNet (Single Shot Multi Box). These NN's have grown in popularity for their application in mobile devices while maintaining a good accuracy at detecting objects. Below is a screenshot of the performance on the RP when using a YOLONet.
The most expensive operation that ShopTrac will have to run is for detecting humans in video frames. This requires using deep learning techniques to identify people. This week, the plan was to study the performance in terms of frame per second (where higher is better) that different ANN's would result in.
Surprisingly, the YOLONet was found to provide poorer performance than the MobileNetSSD. Despite the fact that the YOLONet promised better performance, in testing this was not the case as per the lower average frames per second when processing a video of people walking through traffic. This could have resulted from the fact that the model that was used was trained to locate more than just people, but detect other inanimate objects which likely required quite a bit of processing. The MobileNet provided far better performance but still not enough. We think some optimizations can still be made to further improve this performance but for now it is clear that the MobileNet is the way to go as detection is required.