This thesis studies GPU accelerated visual odometry in measuring log length. The visual odometry would not suffer slippage nor require recalibration depending type of wood or temperature conditions compared to mechanical measurement. The requirement of the real-time performance is quite high. Image capturing in 120 Hz frequency is needed as log is moved several meters per second by harvester heads. Here GPU acceleration will be used as it can give speedup in magnitude of hundreds or more. Real-time performance is targeted by selecting fast algorithms for subtasks of measurement pipeline and considering possibilities to parallelize algorithm. In many cases performance boost is achieved, but not in expected magnitude. Physical constraints of the graphics card hardware become easily the limiting factor in parallelization. Real-time performance was achieved in this thesis but not with required accuracy. It remained for future work to find out which algorithms would give both targets.
|Publication status||Published - 2015|
|MoE publication type||G3 Licentiate thesis|
- Visual odometry
- Measuring methods
- Log length