Rohan Gupta

Problem: how does accuracy of end-to-end steering prediction of autonomous vehicles change when combining image and CAN bus data?

Started Mar. 22, 2022

Abstract or project description

In recent years, end to end steering prediction has grown into a large area of research. The main method of accomplishing end-to-end steering was to use computer vision models on a live feed of video data. In addition, to increase accuracy, many companies added data from a LiDAR and or radar sensor through use of sensor fusion. However, this comes at a large detriment to the financial cost. In this paper, we address this issue by instead adding CAN bus data through the use of sensor fusion to improve the accuracy when compared to a sole computer vision while keeping the cost relatively similar. To collect the data, we drove a car around with a dashboard mounted camera and a computer hooked up to the OBD-II port. We then trained resnet 50 on the video data and CAN bus data and compared this model to one trained without CAN bus data. The metric we choose for measuring success in our experiment is through the Mean Absolute Error (MAE). When we trained the model without CAN bus data we obtained an MAE of x.xx while the model trained with the CAN bus we got an MAE of y.yy. We were able to break which parts of the CAN bus data had the most effect on the accuracy. We did find that the air pressure had an MAE of z.zz which was intriguing since it was very close to zero. As our results are high with the CAN bus data included, it shows that CAN bus data can improve the accuracy while keeping the same financial costs.

Emotion Detection in audio data

Started Apr. 6, 2021

Abstract or project description

This paper will try to predict emotions in audio data. The best accuracy achieved was 82%.