See-Through ADAS

Academic Sep 2017 - Present

See-Through ADAS honors project is an autonomous driver assistance system with stitched video feeds from the perspective of two cars, one behind the other, with useful annotated information. The goal of the this project is to improve situational awareness and efficiency of drivers, so as to reduce driver mistakes on the roads. Our design will accomplish this with object recognition, video stitching, and video annotations.


I devised an object-tracking algorithm with 95% accuracy in a 30 minute test drive by utilizing and training a convolutional neural net with a mAP score of 74.7.


I implemented a linear-time greedy algorithm to minimize the delay between two camera feeds to 5 milliseconds.


Our object recognition is done by a convolutional neural network model (YOLOv3), and based on its results, we annotate the frames using the OpenCV library on Python. The frame stitching is done using OpenCV image processing and computer vision APIs, to perform Laplacian pyramid blending.

screenshot

See-Through ADAS

Academic Sep 2017 - Present

See-Through ADAS honors project is an autonomous driver assistance system with stitched video feeds from the perspective of two cars, one behind the other, with useful annotated information. The goal of the this project is to improve situational awareness and efficiency of drivers, so as to reduce driver mistakes on the roads. Our design will accomplish this with object recognition, video stitching, and video annotations.


I devised an object-tracking algorithm with 95% accuracy in a 30 minute test drive by utilizing and training a convolutional neural net with a mAP score of 74.7.


I implemented a linear-time greedy algorithm to minimize the delay between two camera feeds to 5 milliseconds.


Our object recognition is done by a convolutional neural network model (YOLOv3), and based on its results, we annotate the frames using the OpenCV library on Python. The frame stitching is done using OpenCV image processing and computer vision APIs, to perform Laplacian pyramid blending.

screenshot