Funny, I built a security camera setup with PoE CCTV cameras, GStreamer and the NVIDIA CUDA element using a Xavier platform. I tried SSDMobileNet and YOLO and found them to be absolutely horrible.
The camera that was pointing down at an angle was the worst. Both models would only identify a dog and a person correctly about 15% of the time (missing me or my partner as I walked by and waved), with an actual object detection about 80% of the time even when there was nothing in its ground truths in-frame!! (usually as desks, beds or chairs, i don't recall exactly but it was furniture - and it was pointed at my empty back lot). It had just as many shadow/sunspot/tree failures as Motion. The other camera at eye level did a great job with cars, but not so much with people's side profiles, only head-on.
It was laughably bad. And I have no intention of training my own models on my datasets because I don't have time to label. I did this in 2018-2019 so I don't know what the state of the art object detection models are like today, maybe they got their shit together for non-canonical angles.
I eventually switched back to full-time recording on a 2 TB HDD and if I need to scan back i can jog the livestream because it saves weeks of data.
I had more luck with YOLOv8. But I still keep the motion-detected archive (generated by DVR-Scan) for some months, and the raw footage for a couple weeks as well.
The camera that was pointing down at an angle was the worst. Both models would only identify a dog and a person correctly about 15% of the time (missing me or my partner as I walked by and waved), with an actual object detection about 80% of the time even when there was nothing in its ground truths in-frame!! (usually as desks, beds or chairs, i don't recall exactly but it was furniture - and it was pointed at my empty back lot). It had just as many shadow/sunspot/tree failures as Motion. The other camera at eye level did a great job with cars, but not so much with people's side profiles, only head-on.
It was laughably bad. And I have no intention of training my own models on my datasets because I don't have time to label. I did this in 2018-2019 so I don't know what the state of the art object detection models are like today, maybe they got their shit together for non-canonical angles.
I eventually switched back to full-time recording on a 2 TB HDD and if I need to scan back i can jog the livestream because it saves weeks of data.