Baraja

Projects

Significantly improved detection of distant and dark objects

Through large data analysis of many captured point-clouds, we found that we were losing power to portions of the optical waveform that did not contain coded information, and as such could not be used for object determination.

I led a small cross-disciplinary team of electrical and system engineers to understand if this was solvable, while managing project risks and timelines with the executive team.

We successfully identified and solved the problem by delivering a new approach to Mach-Zehnder calibration, which meant we were able to improve the system's link budget by 3-4dB. Conventional 1/r^2 loss dictates this can increase the range of the LiDAR by ~1.5x. Or, increase the points on a dark target from 10% to ~85%. Both of these improvements were evident in the product.

Further, I designed and rolled out a python data analysis package that enabled manufacturing to test for this fault. This enabled us to pass-fail devices early in the device build, which saved significant engineering and manufacturing effort.

Reduced the system false alarm rate by 100x

Complex devices like LiDARs are prone to many types of noise and interference; some stationary and predictable, some not so stationary and not predictable... but some are not so stationary and still predictable!

My first major impact on the product was to deliver an adaptive noise estimation strategy, which was used to establish a dynamic detection threshold for the LiDAR based on the local noise statistics. Since thermal effects and optical reflections caused the noise floor of the sensor to change dynamically, the estimation of this led to a huge improvement in the ability to discriminate between interference, noise, and the real target.

This meant we were able to reduce the system false alarm rate significantly without degrading our link budget.

Coherent cancellation strategy for self-interference

2022 saw us focus our efforts on a new coherent LIDAR architecture, which can produce both a per-point distance and a per-point velocity estimate with doppler.

Unfortunately, laser phase noise - an unpredictable rotation of the phase of the laser with time caused by spontaneous emission - makes extracting this information exceedingly difficult, especially on dark objects and in the presence of interfering signals such as back reflections.

In spite of the complications of the laser phase noise, I worked with the team to deliver a solution to estimate the contribution of interfering sources, and mitigate its impact on the system by up to 30dB.

This work also involved prototyping and recommending important changes to the device's optical architecture to achieve a solution that worked in the product.

Led the design of from-scratch DSP chains for test, CI and scale

When I joined Baraja, my first efforts were directed to building an integrated DSP chain in a python packaged, version controlled architecture. This meant accurately interpreting the teams expansive list of design reviews and implementing the algorithms in an optimised fashion.

This body of work enabled the team to:

Efficiently process raw-data captures of the point-clouds to easily debug product-facing issues.
Work from a common, version-controlled codebase using Git Flow. The team were able to stop using fragmented and mostly conflicting local versions of the DSP chain.
Adopt a principled "design for test" software development philosophy, enabling our FPGA team to design a coherent and bit-exact test strategy around the code-base.

This ultimately gave the whole business strong confidence that, after investing in large bodies of FPGA implementation work, the product would have a significantly higher probability of matching the algorithm specification.