The Halo Auto Aim Bot is a python file consisting of methods for screen capture, object detection inference, and mouse movement.
During code execution, we first capture the screen on which Halo is being played. From here, the image is sent into an object detection model for inference. The output of this inference is a list of bounding boxes with enemy detections. From here, we loop through all of the predictions to find the object closest to the current aim point. Once this is determined, we calculate the relative distance to the target and define input requiremetns via PID controller and command mouse movement. Optionally, we allow the user to toggle the ability to automatically fire the equipped weapon when the target has been acquired.
Model Selection:
Due to the fast paced nature of the game, any model that we select needs to be able to return accurate results quickly during dynamic gameplay. Additionally, since the game can be graphically demanding, we need to select a model that does not use too much of the GPU’s resources. Three models were selected and tested, with YOLOv5 being selected for final use.
● You Only Look Once (YOLO) v1[4]: YOLOv1 is a single shot object detection model which is computationally light, but has a relatively low precision and recall compared to other state of the art object detection models.
● Faster RCNN[5]: FRCNN is a two stage object detection model which is also relatively computationally light, though more so than YOLOv1. This model is capable of higher precision and recall than single stage models, but is slower in doing so.
● YOLOv5[3]: YOLOv5 is the latest version in the YOLO model family. It is also a single stage detector, but is capable of higher precision and recall than YOLOv1, while at the same time maintaining its ability to perform very fast inference.
docker build -t autoaim .
docker run --rm -d -itp 8888:8888 -v %cd%:/app --gpus all autoaim