February 14, 2018
Rowers face dangers with their back, essentially rowing blind, which can pose a threat to their safety. Therefore, they constantly have to turn around to check for obstacles ahead, leading to exhaustion.
To address these issues, the aim was to equip rowers with an extensive collision avoidance system, allowing them to prevent boat damage without the need to constantly turn around and check their surroundings.
Rowing on a lake or river might pose a great danger to a rower with the vastness and freedom available. But there is one element that is not playing in favor of for a rower and his security: they face danger with their back. Essentially they are rowing blind, completely oblivious to any obstacles ahead. So they turn around every moment to check their surroundings. Let’s try: look behind you and back to your screen. Again. Again. Again. Getting tired, right? Well, Rowcus aims to sink this trouble deep into the sea. They provide rowers with a collision avoidance systems to prevent any damage to their vessels [ like this one ]. Therefore, the rower can focus on rowing and the expensive boat is save.
As usual, the team started with a Design Thinking Workshop to understand the problem before designing a solution. What sensors would be needed? What problems should the solution solve? Which technology was best fit to do the job?
Fortunately for the team, Rowcus was able to provide them with multiple sensors that were employed in similar applications. The goal was to establish the usability of such devices considering the challenges of rowing and its environment ( on water, movement, the variety of obstacles, …).
The problem was first without any constraint (costs, size, feasibility …) and brought up a variety of ideas using all kinds of technology. From fleets of mini-boats to patrol the waters to training an animal to provide the same service. But ultimately, the team started boiling the ideas to their core components to identify the main ideas behind them. Whether it was training a dog to bark when approaching a buoy or implementing the same on computer vision hardware, it all was about training an object recognition system. And so resulted from this late-night DTW a list of technologies and implementation ordered by priority to implement and validate starting with a redundant system around LiDAR and Object Recognition. A simple minicomputer such as the raspberry pi should be enough to do the job of data processing.
The next morning, the work was distributed as to efficiently advance to our multiple goals in parallel. First, the resolution of the PiCamera was surprising and seemed like a good choice at first. In order to test this, the team grabbed a piece of the previously recorded footage by the Rowcus team on the water. Simple research yielded some promising methods for an application as in this case. The recipe was ready and the solution ready to be cooked. Use state of the art technologies and methods of computer vision by OpenCV. Sprinkle it with existing models of objects most likely to be encountered by a rower. Tada! you have an object recognition system ready to be devoured by your GPU. This system immediately proved its power and worth on the previously recorded footage. It delivered all the necessary information: What is the object? Where is it?.
On the other hand, the LiDAR system was causing trouble after trouble. The work on this module turned this breeding weekend into a man-vs-machine weekend, slowing down some of the progress and ever increasing the confusion. It took the random visit of a fellow breeder and a moment of combined epiphany to resolve the issues on the spot. Some lines of codes and a few anger management sessions later, the LiDAR module starting spitting out what seemed like usable data. It was time to combine both systems and use each one to validate the other. Time was running short but a few plots generated on the fly brought some relief to the team.
After lots of head-scratching and deep looks into the void of an empty whiteboard, the team had set up a couple of implementations. A neural network, trained on some obstacles for a rower (boats, shores, …), based on the OpenCV framework could detect and identify objects on the fly. Unfortunately, lacking the compiled framework for the Raspberry Pi and an Earth day having only 23h56m, the live demonstration was not ready for the delivery. The LiDAR Module in the end delivered on some expectations and the team noted, that coupling these systems into a live running system would make it an incredibly powerful tool.
Concluding this challenging and unique weekend, we would like to thank Rowcus for their trust in Unicorn Labs. We wish them all the best in the future in keeping rowers and their boats safe and sound.