Object detection from a vertical position?

[Pentagano]

I've recently installed a camera quite high up and have deepstack with the AI tool running.

Does the object detection api need an angle to work effectively?

I've noticed these objects are not being detected from above
All settings on maximum sensitivity.

Possibly need to lower the cam so it's at a 45 degree angle?


Thanks

 

[sebastiantombs]

Generally speaking, to get a reliable shot for identification purposes the camera needs to be no higher than 7.5 feet from grade level. From the angle you're trying to use, I don't think that a target can be reliably identified by any AI tool as being a human, let alone be able to identify what the look like.

 

[brianegge]

From the angle you're trying to use, I don't think that a target can be reliably identified by any AI tool as being a human

AI can reliably classify objects from a top down view, as is often done from aerial imaging. For my own driveway, I have one camera top down “overview” and then a pair of secondary cameras under the garage eaves. I’ve trained my model to identify people and animals at various angles.

 

[TonyR]

For my own driveway, I have one camera top down “overview” and then a pair of secondary cameras under the garage eaves....

Is it safe to assume this is not a shot of your driveway?

 

[brianegge]

Is it safe to assume this is not a shot of your driveway?

Here's my driveway:

I haven't setup deepstack yet, but I could train a model for it. My setup and training are optimized for small objects, while I think DeepStack is better optimized for speed. Some systems can categorize people based on their uniform, so you can get an alert when a delivery driver is present.

 

[sebastiantombs]

Identifying a human as a human is one thing. Identifying exactly who that human is, is a whole different thing. A vertical view will never be able to do that, AI or not. Being able to know who it is is the real key to video surveillance IMHO. Anything else is just playing with technology.

 

[brianegge]

A vertical view will never be able to do that, AI or not.

I would not be so confident, gait recognition is one technique to do so. Your local law enforcement probably doesn't have a gait recognition database yet, so it may be of limited use for homeowners.

"its system can identify people from up to 50 meters (165 feet) away, even with their back turned or face covered."

Chinese 'gait recognition' tech IDs people by how they walk

BEIJING (AP) — Chinese authorities have begun deploying a new surveillance tool: "gait recognition" software that uses people's body shapes and how they walk to identify them, even when their faces are hidden from cameras.

apnews.com

For notifying when someone is on your property I find the overhead is just fine. If being able to identify someone is important, lower cameras are needed for now.

 

[sebastiantombs]

I still maintain that without a full facial shot at good to excellent resolution, a positive ID cannot be made that will stand up in court. Defense attorneys get rich on things like this. On the other hand, it is getting downright scary the way AI is being used for surveillance purposes.

 

[Pentagano]

I was testing with more human detections and my angle was just too steep to detect in the dark especially.

Moved the camera much further down and in a different position. So far so good.

 

[Pentagano]

and boom got the little critter (or rather big one!)
Triggered the sprinklers and did not step a foot on the lawn to do his business.

 

[sapper]

I haven't setup deepstack yet, but I could train a model for it. My setup and training are optimized for small objects, while I think DeepStack is better optimized for speed. Some systems can categorize people based on their uniform, so you can get an alert when a delivery driver is present.

Are you able to upload a deepstack model for detecting people overhead? I'm playing around with tracking people on a stage in a studio, and an overhead view would really make things simple. The overhead view is less cluttered and tracking individuals much easier (unless they climb over top of each other)

 

[brianegge]

Not sure when I'll get around to training other models, but if someone else is interested my dataset is here: ipcams2 Object Detection Dataset by Egge Public

 

[actran]

Not sure when I'll get around to training other models, but if someone else is interested my dataset is here: ipcams2 Object Detection Dataset by Egge Public

@brianegge Thanks for sharing your dataset. I'm more interested in your approach and best practices to get a good custom model.

Questions:
#1: What scripts or tooling do you use to gather the desired images from your ip cameras?

#2: Did you use roboflow to train your custom models? If so, why did you chose roboflow versus other approaches?