Independent mobility of visually impaired people is key to making an inclusive society for them. Unstructured infrastructure in developing countries pose significant challenges in developing aids to address the mobility problem of visually impaired. Most of the assistive devices available internationally assume a structured and controlled environment severely restricting the applicability of such devices. In this paper, we assess the ability of state-of-the-art assistive devices for addressing the independent outdoor mobility needs of the visually impaired in an unstructured environment. We have created realistic datasets for various scenarios and evaluate deep neural networks for object detection on these datasets. We also present a portable prototype for the task. Further, we have also developed a cloud based solution to address the mobility requirements. We compare the local device based and cloud based solutions in terms of accuracy, latency, and energy. We present and discuss results from these two implementations that can provide insights for an effective solution. The results and insights open up novel research problems for embedded systems. © 2019 IEEE.