Video Surveillance with YOLO+llava

by psychip • 2 years ago

261 65 2 years ago

_giorgio_4• 2 years ago

All I see, usually, is some AI YOLO algorithm applied to an offline video.

This is the first time that I've seen a "complete" setup. Any info to learn more on applying YOLO and similar models to real time streams (whatever the format)?

hug• 2 years ago

This repository seems to be exactly what you are asking for. It's YOLO analysis of video frames passed in through Real Time Streaming Protocol.

_giorgio_4• 2 years ago

Yes, probably it's only one reasonably sized, let's say that with a lot of patience you can study it! I'll search for some online resources too.

I thought that this topic yolo object recognition would have much more following, instead there are really only a few projects.

https://github.com/search?q=yolo+rtsp&type=repositories&s=fo...

llm_trw3• 2 years ago

Just stream it one frame at a time to the model and eat the latency: https://www.youtube.com/watch?v=IHbJcOex6dk if you need more hand holding.

There's a reason why there's a whole family of models from tiny to huge.

yeldarb2• 2 years ago

If you do it naively your video frames will buffer waiting to be consumed causing a memory leak and eventual crash (or quick crash if you’re running on a device with constrained resources).

You really need to have a thread consuming the frames and feeding them to a worker that can run on its own clock.

_giorgio_4• 2 years ago

Thanks for the link, but what happens when you have a video stream, be it a usb webcam, or an RTSP stream, and the hardware can't keep up?

I'm on windows.

Ideally I'd like the frames to be dropped, so the inference is done on the last received frame? Is this a standard behaviour?

yeldarb2• 2 years ago

We’ve got an open source pipeline as part of inference[1] that handles the nuances (multithreading, batching, syncing, reconnecting) of running multiple real time streams (pass in an array of RTSP urls) for CV models like YOLO: https://blog.roboflow.com/vision-models-multiple-streams/

[1] https://github.com/roboflow/inference

ferar• 2 years ago

Can you specify ideal hardware (camera, computer) to deploy the solution? Thanks

llm_trw3• 2 years ago

Default yolo models are stuck at 640x640, so literally any camera that is at least capable of that resolution. Llava I believe is about the same. You'd need ubuntu and something that can run a llava model in vaguely real time, so a 4090/4080.

moandcompany6• 2 years ago

You'll want to find an IP Camera that supports the RTSP protocol, which is most of them.

If your budget supports commercial style or commercial grade cameras, looking at Dahua or Hikvision manufactured cameras would be a good starting point to get an idea of specs, features, and cost.

meow_catrix2• 2 years ago

Maybe don’t buy surveillance hardware from those brands

skirmish• 2 years ago

Here are hardware recommendations from another similar (and well established) project: [1] [2]. Even though they don't recommend Reolink cameras, I have both Amcrest and Reolink cameras working well with Frigate for more than a year now.

[1] https://docs.frigate.video/frigate/hardware

[2] https://github.com/blakeblackshear/frigate

jamesbfb• 2 years ago

+1 for Frigate and Reolink. I have it running in a Proxmox VM on an old dell r710 (yes, it’s sucks watts and needs replacing) but all said, Frigate, is, amazing! The ease of integration with HA is equally great.

moandcompany6• 2 years ago

Many Amcrest IP Cameras are manufactured by Dahua and use localized versions of Dahua firmware. The same applies to the Lorex brand in the United States.

Some things that matter when it comes to configuring your IP Cameras (Beyond security, etc): - Support for RTSP - Configurable Encoding Settings (e.g. h264 coded, bitrate, i-frame intervals, framerate) - Support for Substreams (i.e. a full-resolution main stream for recording, and at least one lower-resolution substream for preview/detection/etc) ...

Make sure the hardware you select is capable of the above.

Configurability will matter because Identification is not the same as Detection (Reference: "DORI" - Detection, Observation, Recognition, and Identification from IEC EN62676-4). If you want to be able to successfully identify objects or entities using your cameras, it will require more care than basic Observation or Detection.

npteljes• 2 years ago

I can recommend the Axis brand. Very user friendly while power user friendly as well, true local offerings. I personally bought mine used, it's an older model, and even then, it holds up really well.

toomuchtodo• 2 years ago

+1 for Axis

rocauc• 2 years ago

A suggestion: I'd swap llava for Florence-2 for your open set text description. Florence-2 seems uniformly more descriptive in its outputs.

jerpint• 2 years ago

I found grounding-dino better than Florence and faster

netdur• 2 years ago

I found YOLOS to be faster and better, bot real time but 22k objects under half second

Eisenstein2• 2 years ago

They are using Ollama which is based on llama.cpp; florence is not supported on that backend.

yu3zhou4• 2 years ago

Congrats! What hardware you use to run the inference 24/7? I built a simpler version for running on low end hardware [0] for recognizing if there’s a person on my parcel, so I know someone have trespassed and I can launch siren, lights etc.

https://github.com/jmaczan/yolov3-tiny-openvino

nikolayasdf123• 2 years ago

how about llama3.2 vision? should it get better performance?

pmontra• 2 years ago

This runs with a Geforce GTX 1060. By a quick search it's 120 W. Maybe it's only the peak power consumption but it's still a lot. Do commercial products, if there are any, consume that much power?

hcfman6• 2 years ago

I have something similar. It's not tracking though. Drawing around 10W on a pi, around 7W on a Jetson.

4ggr0• 2 years ago

not sure if i'm misunderstanding - you've got a similar GPU to a 1060 hooked up to a pi?

formerly_proven3• 2 years ago

YOLO is quick enough that you can just run it on a CPU, assuming you don’t want to run it at full resolution (no point) and full frame rate (ditto) for multiple streams. When you run it scaled down at a 2-3 fps you’ll get several streams per CPU core no problem. Energy use can be minimized by running a quick motion detection pass before, but that would obviously make the system miss things creeping through the frame pixel by pixel (very unlikely if you ask me)

phito2• 2 years ago

You can use a Coral USB Accelerator, doesn't use more than 10W.

nicholasjarnold• 2 years ago

Can confirm. The Coral inference accelerator is quite performant with very low power draw. Once I figured out some passthrough and config issues I was able to run Frigate in an LXC container on Proxmox using Coral USB for inference. It's been working reliably 24/7 for months now.

Eisenstein2• 2 years ago

You can see here:

                res = rest(ollama, {

                    "model": "llava",

                    "prompt": genprompt(box.name),

                    "images": [box.export()],

                    "stream": False

                })

They are calling the ollama API to run Llava. Llava is a combo of an LLM base model and + vision projector (clip or ViT), and is usually around 4 - 8GB. Since every token generated needs access to all of the model weights, you would have to send 4 - 8 GB through USB with the Coral. Even at a generous 10gbit/s that is 8GB / 1.25GB = 6.4seconds per token. A 150 (short paragraph) generation would be 16minutes.

hcfman6• 2 years ago

Yeah. But it’s likely it’s an 8-bit quantised, likely very small model with a small number of parameters. Which translates into poor recall and lots of false positives.

How many parameters is the model you are using with hailo? And what’s the quantisation and what model is it actually ?

moandcompany6• 2 years ago

There's a wide range of inference accelerators in commercial use.

For "edge" or embedded applications, an accelerator such as the Google Coral Edge TPU is a useful reference point where it is capable of up to 4 Trillion Operations per Second (4 TOPS), with up to 2 Watts of power consumption (2 TOPS/W), however the accelerator is limited to INT8 operations. It also has around 8 MB of memory for model storage.

Meanwhile a general purpose or gaming GPU can support a wider range of instructions, single-precision, double-precision floating point, integer, etc).

Geforce GTX 1060 for example: 4.375 TFLOPS (FP32) @ 120W (https://www.techpowerup.com/gpu-specs/geforce-gtx-1060-6-gb....)

There are commercial-oriented products that are optimized for particular operations and precision.

Here's a blog post discussing Google's 1st-generation ASIC TPU used in its datacenters: https://cloud.google.com/blog/products/ai-machine-learning/a...

(92 TOPS @ 700 Mhz - 40W)

https://arxiv.org/abs/1704.04760

janalsncm• 2 years ago

Sorry I’m not familiar with TPUs only GPUs but how much VRAM do Corals have? YOLO 11x is 56M params which if it was quantized to int8 would still be 56MB. Plus you would need some for your inputs.

vaylian• 2 years ago

Hello from the privacy crowd! Please use this responsibly. Tech can be a lot of fun and I encourage you to play around with things and I appreciate it when you push the boundaries of what is technically feasible. But please be mindful that surveillance tech can also be used to oppress people and infringe on their freedoms. Use tech for good!

doctorhandshake• 2 years ago

> It calculates the center of every detection box, pinpoint on screen and gives 16px tolerance on all directions. Script tries to find closest object as fallback and creates a new object in memory in last resort. You can observe persistent objects in /elements folder

I’ve never implemented this kind of object persistence algo - is this a good approach? Seems naive but maybe that’s just because it’s simple.

xrd• 2 years ago

I'm confused about why you need yolo and llava. Can't you simply use yolo without a multimodal LLM? What does that add? You can use yolo to detect and grab screen coordinates on its own, right?

michaelt• 2 years ago

Almost certainly using yolo to segment the cars, then llava for the more detailed "silver sedan" description

andblac• 2 years ago

Skimming through the source it seems to run 'car' and 'person' objects through llava with the following prompt:

- "person": "get gender and age of this person in 5 words or less",

- "car": "get body type and color of this car in 5 words or less".

So YOLO gives the bounding box and rough category, while llava describes the object in more details.

anshumankmr• 2 years ago

Could try with Florence by Microsoft instead of Yolo and Llava, though the results are not going to be as great. Florence will do the inference on CPU. This is just for fun.

011000112• 2 years ago

If you're interested in DIY security+AI, check out Frigate NVR(https://frigate.video/), Scrypted(https://www.scrypted.app/) and Viseron(https://viseron.netlify.app/).

gh02t• 2 years ago

I've been using Frigate for a long time and it's a really cool project that has been quite reliable. The configuration can be a little bit of a headache to learn, but it gets better with every release.

Viserion is new to me though, that looks really cool.

dfc• 2 years ago

I just recently got frigate up and running. How do the other two compare?

011000112• 2 years ago

Beats me, I'm just getting into this now. I started with a Reolink NVR, but it's a piece of crap, so I'm looking for a better alternative.

It looks like either Frigate or Viseron will do what I want. I started setting up Frigate, but realized I should downgrade my Reolink Duo 3 to a Duo 2 before I go too far. The Duo 3 really doesn't offer much better image quality but forces you to use h265 and consumes a lot more bandwidth. Once I stabilize my camera setup I'll get back to setting up both Frigate and Viseron and see what performs better. I like that the pro upgrade of Frigate allows you to customize the model and may make use of that.

taikon• 2 years ago

I've been running frigate for a while now and I find it's object detection has a higher than preferred false-positive rate.

For instance, it kept thinking the tree in my back yard is a person. I find it hilarious that it often assigns a higher likelihood the tree is a person than me! I've needed to put a mask over the tree as a last resort.

acidburnNSA• 2 years ago

Assuming the tree is big you can set max object areas for person and then it will never happen again. I had to do this with some areas where shadows looked like people in the afternoons.

matrik• 2 years ago

MobileNetV3 and EfficientDet are othwr possible alternatives to YOLO. I was able to get higher than 1.5 FPS on Raspberry Pi Zero 2W which draws 1W on average. With efficient queuing approach, one can eliminate all bottlenecks.

Video Surveillance with YOLO+llava

Related Stories