Time-to-Collision Algorithm and Real-Time Implementation

The time-to-collision task is that of estimating the ratio of the distance of an object to the relative velocity in the direction of that object, based on time-varying images resulting from the 3-D motion of a camera with respect to the object. Traditional methods require large amounts of memory and computation time, and are not suited for real time hardware implementations. We used a direct method to provide a non-linear closed form solution for the case of translational motion, because it is more robust to noise, and provides a straightforward solution for time-to-collision. Three approaches gradient descent, a grid search method and an improved grid search were used for solving the non-linear equations. Two alternatives analog VLSI chip implementation and digital signal processing microprocessor based system design were investigated. Finally, the time-to-collision algorithms were simulated on a Texas Instruments digital signal processor(DSP) TMS320C67 based system. From the results of the simulation on the TMS320C67 Code Generator and Simulator, it is shown that the improved grid search method provides reliable results and high computational efficiency with real time performance when implemented with the DSP TMS320C67. Thesis Supervisor: Berthold K.P. Horn Title: Professor of Electrical Engineering and Computer Science