YOLOv4 with TensorFlow 2

Posted October 06, 2020 by Rokas Balsys

Counter-strike Global Offensive realtime YOLOv4 Object Detection aimbot

Welcome to my last YOLOv4 custom object detection tutorial, from the whole series. After giving you a lot of explanations about YOLO I decided to create something fun and interesting, which would be a Counter-Strike Global Offensive game aimbot. What is aimbot? Actually idea is to create an enemy detector, aim at that enemy and shoot it. The same idea we could use in real life for some kind of security system, for example, detect and recognize people around our private house and if it's an unknown person aim lights at that person or turn on some kind of alarm. Also, I thought that it would be fun to create a water gun robot, that could aim to an approaching people and shoot the water, why water? To avoid the violence of course.

But I am not talking about what we could do, lets talk about what I did. So, first I took my YOLOv4 GitHub code, on which I was working half a year to make it easily understandable and reusable, I created a lot of tutorials to explain every part of it.

Then I used the technique explained in my previous tutorial to generate training data for my model to make it quite accurate while detecting enemies.

First, I would like to discuss instructions, how you can run it by yourself, then about results and lastly, I will talk about code, so there will be 3 parts.

How to run/test GitHub code:

My testing environment:

  • i7–7700k CPU and Nvidia 1080TI GPU
  • OS Ubuntu 18.04
  • CUDA 10.1
  • cuDNN v7.6.5
  • TensorRT-
  • Tensorflow-GPU 2.3.1
  • The code was tested on Ubuntu and Windows 10 (TensorRT not supported officially)

Why ubuntu? Because it's much easier to run TensorRT on Ubuntu than Windows 10, you can try to run it on Windows 10 with the following repository.

First, you must install TensorFlow, Python 3, Cuda, Cudnn and etc packages to prepare the TensorFlow environment. Second, if you are on windows it's quite easy and obvious how to install Steam. But if you are on Linux, it's a little harder. I used a flatpak to do that:
1. sudo swupd bundle-list | grep desktop
2. sudo swupd bundle-add desktop
3. flatpak install flathub com.valvesoftware.Steam
Then run steam with the following command in terminal
4. flatpak run com.valvesoftware.Steam
There still may be errors while installing or running steam, but you should solve them with google help :).
5. In steam download Counter-Strike Global Offensive

When you have steam and CSGO downloaded, we can download my GitHub repository. You can clone it or download it as a zip file, it doesn't matter. I already zipped my trained model, which I put into the checkpoints folder. If you are on Windows, unzip it using 7zip, if you are on Linux there are two ways how to unzip:

1. Download P7ZIP with GUI and unzip everything.
2. Install required packages
sudo apt-get install unzip unrar p7zip-full
ython3 -m pip install patool
python3 -m pip install pyunpack. Now open the checkpoints folder and run linux_unzip_files.py script.

Now you need to install all requirements:
pip install -r ./requirements.txt

Now, everything is ready. My yolov3/configs.py file is already configured for custom trained object detection with input_size of 416. We simply need to run it. You may change YOLO_INPUT_SIZE if you need better accuracy, but you will lose in FPS. Now, when you have running the CSGO game in the background, simply run YOLO_aimbot_main.py script. When YOLO detects objects on the screen, it should start moving the mouse and shooting the enemies.

If the mouse is flying around in-game, open the game console, and type m_rawinput 0 this will disable raw game input. Also, you may need to change sensitivity or other minor settings.

Achieved results:

So here is a short GIF from my results, check my YouTube tutorial for more.


First, I should tell you that I used only around 1500 images to train my aimbot model. Most of this training data I generated with the method I explained in my previous tutorial. To make it even more accurate it's recommended to use more than 10 thousand images in different maps and so on, then we would be sure that our model won't detect enemies wrong. Best to understand my results would be to watch my YouTube video. Anyway, I ran 3 different test instances:

  • TensorFlow detection with 416 input size:

  • TensorRT INT8 detection with 416 input size:

  • TensorRT INT8 detection with 608 input size:

So, what these Frames Per Second results tell us? At first, I used standard YOLO TensorFlow detection, without TensorRT optimization. This is what you can get on Windows 10 with 1080TI GPU, but if you have a newer GPU, you can get better results.

Then I converted my TensorFlow model to the TensorRT INT8 model with an input size of 416. As you can see FPS increased more than double, that's what I was talking about. Mostly I would use this model for small maps, where our enemies come closer to us because it's not that accurate with small objects.

And the last one is TensorRT INT8 model with an input size of 608. As you can see FPS is not that great, but I am sure that accuracy is very high. I would like to have NVIDIA 3080 or even 3090 to see what I could get with it. So, I am sure that the latest generation video card could reach much better results, so even if my current results are not so bad, can you imagine what you would get with these new cards?

A little about the code:

There is nothing a lot to talk about the code, if you have experience in Python, it will be quite easy to understand the main script. For this tutorial, I wrote YOLO_aimbot_main.py script which you can find on my GitHub repository.

def getwindowgeometry():
    while True:
        output = subprocess.getstatusoutput(f'xdotool search --name Counter-Strike getwindowgeometry')
        if output[0] == 0:
            t1 = time.time()
            LIST = output[1].split("\n")
            Window = LIST[0][7:]
            Position = LIST[1][12:-12]
            x, y = Position.split(",")
            x, y = int(x), int(y)
            screen = LIST[1][-2]
            Geometry =  LIST[2][12:]
            w, h = Geometry.split("x")
            w, h = int(w), int(h)
            outputFocus = subprocess.getstatusoutput(f'xdotool getwindowfocus')[1]
            if outputFocus == Window:
                return x, y, w, h
                print("Waiting for window")

If you would like to use this script on Windows 10, you would need to modify it a little. Because I wrote a specific getwindowgeometry() function, that I use to get the geometry of my Counter-Strike window, I mean to get coordinates of the window. Also, if this function can't find the Counter-Strike window, it keeps searching and doesn't run the whole aimbot loop. Of course, you can implement the same stuff on windows, it's even easier, but because I don't have windows 10, I skipped this part. Or you can simply use X, Y, W, H coordinates instead, but check if they are correct at first.

And here is the main while loop:

while True:
    t1 = time.time()
    img = np.array(sct.grab({"top": y-30, "left": x, "width": w, "height": h, "mon": -1}))
    img = cv2.cvtColor(img, cv2.COLOR_RGBA2RGB)
    image, detection_list, bboxes = detect_enemy(yolo, np.copy(img), input_size=YOLO_INPUT_SIZE, CLASSES=TRAIN_CLASSES, rectangle_colors=(255,0,0))
    cv2.circle(image,(int(w/2),int(h/2)), 3, (255,255,255), -1) # center of weapon sight

    th_list, t_list = [], []
    for detection in detection_list:
        diff_x = (int(w/2) - int(detection[1]))*-1
        diff_y = (int(h/2) - int(detection[2]))*-1
        if detection[0] == "th":
            th_list += [diff_x, diff_y]
        elif detection[0] == "t":
            t_list += [diff_x, diff_y]

    if len(th_list)>0:
        new = min(th_list[::2], key=abs)
        index = th_list.index(new)
        pyautogui.move(th_list[index], th_list[index+1])
        if abs(th_list[index])<12:
    elif len(t_list)>0:
        new = min(t_list[::2], key=abs)
        index = t_list.index(new)
        pyautogui.move(t_list[index], t_list[index+1])
        if abs(t_list[index])<12:

    t2 = time.time()
    times = times[-50:]
    ms = sum(times)/len(times)*1000
    fps = 1000 / ms
    print("FPS", fps)
    image = cv2.putText(image, "Time: {:.1f}FPS".format(fps), (0, 30), cv2.FONT_HERSHEY_COMPLEX_SMALL, 1, (0, 0, 255), 2)
    #cv2.imshow("OpenCV/Numpy normal", image)
    #if cv2.waitKey(25) & 0xFF == ord("q"):

Everything is quite easy. First, we grab a part of our game screen, and we use our YOLO detection to detect all objects on that image. As a return, we receive an image with detected objects, detection_list (used to sort detections), and bboxes.

With cv2.circle line: cv2.circle(image,(int(w/2),int(h/2)), 3, (255,255,255), -1) we draw a white dot at the center of our weapon sight, I used it to debug the targeting process.

With for detection in detection_list for loop, I sort all the detections. Here my main goal is to find th-(terrorist head) and t-(terrorists) and put their center coordinates into two different lists. Actually, this is not a center coordinates, but distance-how far the enemy is from our weapon sight in pixels.

Next, my main goal is targeting to a head, because it's much easier to shoot an enemy, although I am making an aimbot here… So, if our detected enemy is not further away than 12 pixels I am starting to shoot that enemy. Otherwise, I am trying to find the closest target to me and aim at it. And I am doing the same process all the time, it's quite simple.

Command cv2.putText(image, "Time: {:.1f}FPS".format(fps), (0, 30), cv2.FONT_HERSHEY_COMPLEX_SMALL, 1, (0, 0, 255), 2) is used to put frames per second text on our detected image, but it doesn't matter if you are not using cv2.imshow line.


So it was quite a nice and interesting journey while creating this YOLO tutorial series for you my friends. This tutorial is only one example of thousands, where we can use object detection. If I would be having the latest generation Nvidia card results would be even more impressive. This tutorial proves, that using machine learning we can automate almost any game, we just need to have resources, time, and knowledge for that.

You can download this project files and use them at your own risk, you can continue developing on it, you can even develop a reinforcement learning agent (I would be really impressed to see one).

I believe that with Nvidia 3090 GPU, with 10k training images, and if you know how you could optimize my code, even more, by improving grabbing screen, detection, and post-processing times, you could achieve amazing results. This could be some kind of an undetectable legit cheat, lol :D

Anyway, thank you all for reading, hope this tutorial was useful for you, like this article, subscribe to my YouTube channel, and see you on the next tutorial!