Inspiration
Last summer, we worked on manually annotating video datasets that featured either e-scooter riders or bicycle riders. Although the task itself wasn't particularly difficult, it became overwhelming with large datasets. A human reviewer had to watch the videos and manually record timestamps, which was both time-consuming and prone to error. This experience highlighted the potential value of an AI-powered classifier to automate this process and save significant time and effort.
What It Does
The concept is straightforward: the trained model takes an image as input and determines whether it depicts a pedestrian, an e-scooter rider, or a bicycle rider.
How I Built It
I utilized PyTorch and transfer learning to fine-tune a ResNet-18 model for image classification.
Challenges We Ran Into
The biggest challenge was training the model. Image classification requires substantial processing power and time. Even with the resources on Kaggle, training and fine-tuning the model took approximately two hours to complete.
What's Next for Pedestrian Classification Using Deep Learning
The current model, while functional, is limited to single-image classification. To expand its utility, additional classifiers could be developed for traffic lights and traffic signs. Integrating these specialized classifiers with a YOLO-based object detection system could enable comprehensive detection and classification of common road objects. This could then be used to automate annotation for automated driving and path prediction data.
Log in or sign up for Devpost to join the conversation.