Computer vision demos related to classification, object detection, and segmentation for commercial and industrial use cases. I have a large number of computer vision benchmarking scripts, demos, experiments, solution ideas and the like that aren't necessarily big enough for a stand-alone project but are still worth sharing; the goal of this repo is to aggregate them all into one place for later use.
General Approach: everything shared here should be "good enough" to be useful or for the next stage of a conversation to be around what needs to built around the demo to deliver a minimal viable product or proof of concept. E.g., it doesn’t just count cars, it can count cars, handle exceptions, has logic to handle specific events and creates a JSON payload for external systems to consume.
TL/DR - this repo will lean heavily towards practical examples of how to use these technologies, and will often include things like database integration, slack integration, monitoring dashboards, etc.
Using smart thread management and asynchronous functions, we're able to run two video streams simultaneously at about 30 FPS with post processing and ~48 FPS/throughput for inferencing. Not bad for a device that fits in the palm of your hand. Note: I've tried this with as many as six streams: three was at mid 20s FPS and 4-5 was around 18-19, six was around 12-14 with a lot of stuttering. I.e., 2-3 streams seems optimal.
- Trying to do screen capture on top of the two streams grinds things to a halt, so recording with my phone was the best option at the moment.
- Adding the latency to draw boxes + render the frame costs about 2-3 FPS, so the real on screen FPS is closer to 27-28.
- Both videos are 640 x 360
Counting people as they move past a certain point or "border line" in a video. E.g., people going up or down in an escalator. The entry/exit data, total number of people in the frame, and FPS is collected into a JSON payload for sharing with/transmitting to other systems.
Note #1: FPS refers to processing speed, not the rendering speed which is ~30 FPS for the original video and around 10 FPS for the gif
Note #2: in a real life implementation the org would have systems/technology in place that display and store video, so we probably wouldn't render/show video with detections, we would instead just make the data available for later view/analysis whether that's storing the data + the video with detections or just storing the data.
Counting entrances and exits for several different things or classes, think cars going by, people, people on bycycles, dogs, etc. Similar to the above, the demo generates a JSON payload with entry/exit data for each class, and there is an alternate version that transmits data via MQTT to be recorded in InfluxDB for display via Grafana.
Note #1: FPS refers to processing speed, not the rendering speed which is ~24 FPS for the original video and around 20 FPS for the gif
Note #2: the dashboard updates every 5 seconds vs the on screen data updating with every frame, so the dashboard lags the events in the gif/video.