Findings from the research confirmed that a wearable technology-based system equipped with a stereo camera running on an embedded system can take advantage of current state of-the-art computer vision and deep learning approach to detect and classify surface discontinuity in an outdoor urban environment. Through the development of a taxonomy, the research also identified nine classes of surface discontinuity relevant to urban navigational pathways which helped in building a machine learning model. A stacked convolutional neural network was found to be accurate and efficient for classifying the surface discontinuities. With specific hardware used in the implementation of the system, the prototype was evaluated for its accuracy and efficiency. The optimized model achieved a classification accuracy of 96% in field evaluation. In the efficiency evaluation, the prototype was analyzed for its CNN model’s algorithmic efficiency, memory usage, speed and power consumption. The model was 50% more efficient with very low memory usage as compared to the original model it was based on. The processing speed was near real time and the power consumption of the prototype is moderate. The research contributed a methodology and a framework consisting of some algorithms, models and a proof of concept in developing the prototype.
The wearable state-of-the-art system proposed in this study should be a major contribution to the mobility needs and limitations of the blind and low vision people. Great job!
As this study might lead to efforts on the development of a "user interface", as well as "human to computer interactions," developing a prototype to assist the BLVs should be in place. Also, a comprehensive understanding of feedback and communication mechanisms among the BLVs should play a vital role to support bigger projects along this field.