Identifying the location of an autonomous car with the help of visual sensors can be a good alternative to traditional approaches like Global Positioning Systems (GPS) which are often inaccurate and absent due to insufficient network coverage. Recent research in deep learning has produced excellent results in different domains leading to the proposition of this thesis which uses deep learning to solve the problem of localization in smart cars with visual data.
Deep Convolutional Neural Networks (CNNs) were used to train models on visual data corresponding to unique locations throughout a geographic location. In order to evaluate the performance of these models, multiple datasets were created from Google Street View as well as manually by driving a golf cart around the campus while collecting GPS tagged frames. The efficacy of the CNN models was also investigated across different weather/light conditions.
Validation accuracies as high as 98% were obtained from some of these models, proving that this novel method has the potential to act as an alternative or aid to traditional GPS based localization methods for cars. The root mean square (RMS) precision of Google Maps is often between 2-10m. However, the precision required for the navigation of self-driving cars is between 2-10cm. Empirically, this precision has been achieved with the help of different error-correction systems on GPS feedback. The proposed method was able to achieve an approximate localization precision of 25 cm without the help of any external error correction system.
Computer Engineering (MS)
Department, Program, or Center
Computer Engineering (KGCOE)
Bag, Suvam, "Deep Learning Localization for Self-driving Cars" (2017). Thesis. Rochester Institute of Technology. Accessed from
RIT – Main Campus