The basic idea is not new - the development of neural networks in the realtime mode allowed to classify objects in a video stream, such as vehicles, people, animals ...
On the Internet you can find a lot of different projects on this topic.
It would be nice if the driver had been warned about potentially dangerous encounters with objects on the road.
So first idea, then the development and implementation ..
Since the classification of objects in the video stream requires a lot of computational resources, the option with the neural network to the CPU immediately dropped, so get sane FPS is very difficult.
Therefore, the choice fell on the GPU, specifically graphics card from Nvidia with the need Cuda.
Geforce gtx 1063 is able to process (classify objects) in the video stream with a frequency of about 20 frames per second.
Geforce RTX 2080ti a rate of about 30 frames per second.
Comparing the value of the cards and the performance you can safely opt for Geforce gtx 1063.
20 frames / sec is enough for comfortable work.

The entire project was created with the help of tools such as Python, yolov3-tf2, cvui and tensorflow-gpu.
work screenshots can be seen here:




So the program is able to detect objects, classify them, and to warn when approaching them, to define the band, and warn when going beyond them.
The program has settings for thinning the danger zones, focus zones, etc.
Well, half the work is done. But how to install the system in a motor vehicle other than a sound alarm, the driver must receive and visual information.
Simple pilikane not give information on the location of the object.
Put the monitor, such as the idea is not the best. Firstly it will distract the driver, forcing him to move his attention from the road to the screen.
In the second installation of additional accessories on the dashboard may violate feng shui))
Any projections on the windshield of course also have disappeared by themselves.
Those. dead end. The program is, and how to display visual information to the driver is not clear.
In general the program has become covered with dust ..
But recently I got the idea, but why not use this address for the LED strip WS2812B ?!
Take a length of tape as a torpedo car. set it on the edge, near the windshield and you're done!
The tape does not interfere with the review, it does not distract the driver, but at the same time allows you to visualize a place where there is a potentially dangerous object.
Ribbon RGB, so you can easily set color classification.
Such as red - this transport, and purple - people and animals.
The width of the emission is equal to the width of the object in the video stream, and the LEDs light up in the place where the facility is located.
Those. if there was a pedestrian on the right, respectively, then the right will light up the LEDs in purple.
The picture shows the tape, where the center is not well lit and the red LEDs on the right purple.



Treated so .. in front of the machine and is already far away, but yet dangerously close, and on the right a pedestrian.
Pedestrians and animals special priority in processing.
To transfer data from the PC onto the tape using Arduino Nano, which is connected to the com port on the PC.

Like as planned progress: a program, there is visualization. The case for small - to set everything in the car.
System components include:
motherboard, processor supporting avx instructions, memory, SSD drive and Geforce gtx 1063, RGB strip and arduino to transfer data from the PC onto the tape.
The total power consumption must not exceed 120W.
For food, I think you can use PicoPSU at 300 W, will not be superfluous.
The most acute problem - this condensate, how to fight it in the winter, the question I do not have an adequate solution.

So far, I see no reason to spread the source code, as long as there is no single concept of the final system.
At this stage it is yet raw.
Actually that such an idea, and here is its implementation at this stage.
If someone has an idea of ​​the struggle with the condensate, I will be glad to hear.
It is also interested in the opinion regards such imaging systems RGB tape.
And opinions with regards to the use of the PC in the car.
Actually comments on tips are welcome.

The prototype is ready.
Video car recorder:https://youtu.be/9k-Gfr08VKo
The video shows the visualization using the address of the tape.
On consumption can say the following:
- used processor is the most energy efficient, which was close at hand with this avh instructions i7 3770K
- 1063 video card with a small daunvoltom
Total consumption of 110W wattmeter.
I think that the new generation of processors and 1050 cards will take consumption of 80W in the area, which is very good for the car.
While waiting for pico dog for PC power is currently used converter 12c to 220 .. not the best solution, but to test on the road completely.

Post has been editedE_tayson - 07.03.20, 00:29

E_tayson, in essence, you are re-invented mersedesovsky night vision systems. It looks exactly the same (on the monitor).

With regards to the LED strip. Personally, my perception - it's a nightmare. Firstly, it is very distracting. Second - this is an additional bright spot in front of the eyes, in the dark blinds. Third, the detection distance is too small, the objects on it, and so has everyone will see. Fourth, lack of false positives and them where necessary. Fifth, if you go with it all the time - you start to focus on it in the first place, rather than your eyes. It is understood that the tape fail sooner or later. If we do an assistant, it should not duplicate what is so obvious - and to do something with the problem. For example - put the system in IR format, and try to build a system of night vision, as it is - a very, very sick theme.

E_tayson,
looked ... good idea, but the tape does not roll - distracting and informative ... there need wide-angle lens in the short throw projectors and spot lights (designation) of the object on the windshield in the 1 \ 3 square in the lower chasti..i response is necessary increase ... here's an idea for a startup ... Thank you is not necessary - just shared the idea;)
and we should probably do in Linux on Malinka podobnoe..na or windsurfing as it is not the ...

Post has been editedalkaid - 07.03.20, 12:09

"Long-range" system depends on the resolution of the video stream - the higher the resolution, the more accurate the neural network will be able to identify objects at long range, but at the expense of processing speed.
We need a sane FPS, ie. task is to find the "golden" middle.
Here is an example of a system with more customization options, but they are not conclusive:https://www.youtube.com/watch?v=PFhMut0rv0
Transport is now considered potentially dangerous only to the car. While pedestrians and animals on the entire frame, but with a different range to the object based on the object position relative to the car.
Besides allowing fine tuning of the system is the right choice aproksimizirovannogo distance to the subject, before it is added to a list of potentially dangerous object.
At the moment, has already installed the photoelectric cell and changes in the firmware tape control unit. Those. the brightness of the tape depends on the exterior lighting.
Afternoon shines brightly, and at night is not dazzled by the word at all.
Regarding "raspberries", this project is not exactly for her t..k power of the GPU graphics card is used.
You can run both on Windows and on Linux, it does not matter. All written in Python.

Post has been editedE_tayson - 14.03.20, 23:47