AutoDrive

Autonomous Driving: Pedestrians and Vehicles Detection using YOLOv11

This project is part of the Data Science and AI Course at Faculty of Electrical Engineering Sarajevo. We utilize the YOLO (You Only Look Once) object detection model to detect pedestrians and vehicles in driving scenarios, classify traffic lights and track lane, primarily using dashcam footage and images. Different configurations and sizes of the YOLO model were explored to evaluate performance.

Overview

Team Members:

[Emin Hadžiabdić] - (GitHub Profile)
[Armin Memišević] - (GitHub Profile)
[Muhamed Pašić] - (GitHub Profile)
[Bakir Kosovac] - (GitHub Profile)

What to expect in AutoDrive v2:

AutoDrive v1 was accurate but too slow, because it is running 3 models simultaneously on one frame and that is not suitable for real-time applications. In this version we are using 2 models, one for detection and one for segmentation.
The detection model is used to detect objects in first frame, while the segmentation model is used to segment the lane and road on second frame. This approach is more efficient and allows for real-time applications.
But still we are not satisfied with using two frames, in future versions we are going to work on merging all models in one, if it is possible since YOLO is strict when it comes to annotation formatting and merging segmentation and detection tasks, and displaying them on one frame while maintaining real-time performance.

Installation

Clone the repository:

git clone https://github.com/bkosovac1/redziBOT

or clone on Replit:

Datasets

Detection Model
The dataset used for training and evaluation of detection model is available on RoboFlow. It consists of approximately 16,346 frames of classes relevant to this project: green-light, red-light, yellow-light, pedestrian, car, truck, bus, motorbike, bicycle.

Segmentation Model
The dataset used for training and evaluation is available on RoboFlow. It consists of approximately 1927 frames capturing roads and side-lines. Classes are lane and road. We used same model in version 1

Training

Training was done on nano and large size of YOLOv11 model. The training process was optimised with recommended optimizer for that specific dataset. Following are the training parameters and results of all models. Training was conducted using various sizes of the YOLOv11 model (nano and large). Key training parameters and validation results are summarized below:

Model	YOLO Size	Epochs	Batch Size	Learning Rate	Optimizer	Momentum	Dropout	Precision (P)	Recall (R)	mAP50	mAP50-95
Detection Model	`large`	`10`	`16`	`0.001429`	`Auto: AdamW`	`0.9`	`0.0`	`81.1%`	`68.4%`	`75.5%`	`44.2%`
Detection Model	`nano`	`25`	`32`	`0.001429`	`Auto: AdamW`	`0.9`	`0.0`	`81.1%`	`65.6%`	`72.9%`	`41.9%`
Segmentation Model	`large`	`50`	`16`	`0.001667`	`Auto: AdamW`	`0.9`	`0.0`	`83.2%`	`86.4%`	`85.8%`	`77.2%`
Segmentation Model	`nano`	`100`	`64`	`0.001667`	`Auto: AdamW`	`0.9`	`0.0`	`82.9%`	`88.9%`	`86.8%`	`78.7%`