Efficient Visual Localization on TDA4VM (Rev. A)

As vehicles and robots move towards higher levels of automation, accurate localization within a map has become vital. However, since localization is one of many algorithms that need to run concurrently on an embedded platform, it is essential that accuracy does not come at the cost of efficiency. In this application report, an implementation of an efficient localization algorithm on a TDA4VM device, that meets these market requirements is described. The TDA4VM System-on-Chip (SoC) is the first commercially available device in the JacintoTM 7 family of SoCs, a family of SoCs designed from the ground up specifically for automotive and similar robotics applications. This family of SoCs comprises of two primary variants, the TDA4x family designed for ADAS and similar perception tasks in robotics, and the DRA8x family designed for cloud-connected gateway systems. The localization algorithm described here is implemented on the TDA4VM device, a device from the TDA4x family, because the feature extraction portion of the algorithm is based on a Deep Neural Network, and the TDA4VM is equipped with one of the industry's most power efficient deep learning engines. Table of