Embedded Files
Team Members: Stephanie Yang, Dana Atassi, Grace Xiao, Josie Lee, Tanvi Modugula, Sean Lee
Faculty/Graduate Students: Sarjana Sachidanandam
I4C Teaching Assistant: Ava Petusky
Project Overview
Our project seeks to investigate AI's potential to predict human movements in crowded scenarios accurately while ensuring safe navigation towards a designated destination.
Project Question
What machine learning approaches can we use to safely simulate navigation through crowded environments?
Deliverables
AI model that can accurately use a machine learning algorithm to track and predict the movements of human beings in crowded environments based on data set eth.
Our Action Steps
Step 1: Data Acquisition
In this step, we will import the necessary libraries and upload the dataset to our environment.
Step 2: Data Analysis
We will thoroughly examine the dataset to gain insights into its structure and content. Any required code additions or updates will be made during this process.
Step 3: Model Training and Testing
With the dataset prepared, we will proceed to train and test our model using appropriate methodologies.
Step 4: Evaluation
After training, we will review the results and assess the model's accuracy and performance.
Machine Learning Algorithms Used
Neural Networks
A Neural Network is a type of classification algorithm (an algorithm to predict the correct label for data inputted). Neural networks, in contrast to other algorithms, contain at least one hidden layer to pass features from an input layer through to get to the output layer.
Convolutional Neural Network (CNN)
Convolutional Neural Networks use ‘convolution’, a mathematical operation used to filter information. In our program, CNN is used to identify and process the movement of humans each labeled uniquely.
GRAPH Neural Network (GNN)
A graph neural network is used to predict and classify graph properties. This is important in our project because it relies on the making connections with edges and nodes. In our case, the people would be the nodes and the lines connecting them would serve as the edges. Therefore, GNN is crucial to our project.
Results
Alongside adjusting the learning rate and epochs, we also adjusted two variables named FORWARDWEIGHT and OBSTACLEWEIGHT which represented the importance placed on moving forward, and the importance of avoiding obstacles respectively. We ended up finding weights that allowed the robot to succeed with no collisions (with other pedestrians).
Eth dataset
Forward weight was 0.9 and the obstacle weight was 6000.
Hotel Dataset
Forward weight was 5 and the obstacle weight was 6000.
outputVideo1.mp4This video showcases our robot successfully navigating though a crowed hotel environment without any collisions.
Picture of our robot having no collisions in the eth dataset.
Page updated
Report abuse