Model-Based Visual Tracking via Maple Code Generation

Abstract

<p>Many algorithms, particularly in the area of image processing, are expensive to develop and computationally resource intensive. We illustrate the advantages of symbolic code generation using an example - closed-loop visual target recognition and tracking in extreme lighting conditions. We quantify the effect of symbolic code generation methods on code efficiency, and explain how these methods allowed us to reduce the development time as well as improve reliability. Working directly with symbolic models improves software quality by reducing transcription errors, and enabled us to rapidly prototype different models for the visual tracking application, where the need to evaluate trackers in their real-time context precludes the effective use of scripting languages. We describe the model in detail, including formulations as an optimization problem; explain the challenges in solving the model; present our method of building the solvers; and summarize the impact on the performance of our methods.</p>