CS 426 Senior Project

Object Detection in Visually Degraded Enviornments

---

Currently, autonomous vehicles have the ability to function in ideal environments, for example on a clear and sunny day. However, in visually degraded environments, these vehicles have a much more difficult time localizing themselves and detecting objects on or adjacent to the roadway. When a car is operating in low-light conditions or in inclement weather, the current computer vision modules that the cars use to navigate have difficulty detecting objects in their surroundings. A solution has yet to be found to successfully navigate these environments in real time without reliance on additional LiDAR and GPS sensors. For our capstone project, we seek to make a contribution to the problem of object detection in dark environments. We will approach this by gathering video data from a vehicle operating at night time and training an object classifier to detect and locate vehicles in these frames. Using PointGrey cameras, our intent is to collect the data using the Robot Operating System (ROS), and then train an object classifier for vehicles in these conditions. We will use the open source Python framework TensorFlow to construct the neural network and train our image classifier.

With the ability for an autonomous vehicle to detect features in these varying conditions, it can become a commercially viable consumer product as it would permit end-users to operate autonomous vehicles in the dark, when it is raining, or when it is snowing. In addition, this feature would provide users with a higher guaranteed level of safety.