February 3

The team is hard at work finishing up CADs and Layouts. We are trying to problem solve through the global chip shortage and find something that will work.

Mechanical Update

The detailed mechanical design of the main module was completed this week. The design built upon the existing layout from the last blog update; adding part divisions for 3D printing, assembly hardware and mounting straps. The first prototype design has been 3D printed and further design improvements are being implemented to improve the assembly process.


Main Module CAD Design

The basic layout of the rear module was also completed this week, with the layout decisions driven by the optimal wiring and cooling paths. The detailed design of the rear module is the next step.


Rear Module Basic Layout

Electrical Update

The review of the schematic design was completed, and minor changes were made to increase efficiency and reduce heat dissipation. All remaining components for the PCB were ordered and received, and PCB layout for the main, left, right, and rear PCBs is underway. Once layout is completed, the PCBs and stencils will be ordered from PCBWay or a similar electronics manufacturer, and assembled in Waterloo.



Software Update

As mentioned in our previous blog post, one of the main challenges that the software team was facing was the selection of a hardware accelerator for object detection. In addition to the front and rear stereo camera modules, our system also includes left and right facing side cameras for vehicle detection in the stereo camera blindspots. Per our mechanical design, both side cameras must interface with a single Raspberry Pi, meaning that one board will be required to process 3 video streams while meeting our 10 Hz processing specification.

The best option would be a Google Coral USB Accelerator yielding inference times as low as 14.9ms. Unfortunately this product is out of stock at all retailers and very challenging to purchase even from third-parties on eBay. We have decided to purchase the Intel Neural Compute Stick 2 instead which offers inference times as low as 52.8ms. While it is unlikely that a single neural compute stick will allow us to satisfy our 10 Hz processing specification, it is possible to use more than one device to achieve higher amounts of acceleration. We expect to receive this device in the coming days and will perform our own set of benchmarks to evaluate its potential.

So build away, we do.

Come back and check out our blog for regular updates.

University of Waterloo
  • Group 9
  • Falcon Safety