Project descriptions

LiDAR point cloud processing algorithms
LiDAR point cloud processing algorithms 1
LiDAR point cloud processing algorithms 2

LiDAR point cloud processing algorithms

 

LiDAR data is critical in sensor fusion algorithms for autonomous driving. Algorithms operating on LiDAR point clouds are a crucial part of the processing performed by the vehicle's on-board computer. Parallelization and optimization of these algorithms in order to meet latency requirements are of prime importance for real-time systems.

Our objectives in this project were the implementation of point cloud pre-processing “kernels” running on Imagination’s GPU platform using OpenCL, comparison against a reference CUDA implementation, visualisation of the processed point cloud and optimisation and testing on the target hardware platform.

Edge AI supports people with impaired vision
Edge AI supports people with impaired vision 1
Edge AI supports people with impaired vision 2

Edge AI supports people with impaired vision

 

Support of people with impaired vision by detecting nearby obstructions has been achieved by applying computer vision with a depth camera integrated into a walking stick.

Our AI-based object detection system recognises surfaces and moving objects based on proximity and the potential hazard and communicates this using a haptic controller and a Bluetooth earpiece.

The first prototype implementation is being trialled with Blackworld, a blind and partially sighted organisation in Poland.

Knowledge Distillation KD for image reconstruction
Knowledge Distillation KD for image reconstruction 1
Knowledge Distillation KD for image reconstruction 2

Knowledge Distillation (KD) for image reconstruction

 

Ray tracing is computationally intensive: ML techniques can be used to improve performance significantly. Our SoC customer’s initial implementation operated on an entire HD image. Division into tiles, which can be processed in parallel, improves performance on existing hardware, allowing lower quality ray tracing images to be enhanced using neural networks.

Identifying structural defects in silicon production
Identifying structural defects in silicon production 1
Identifying structural defects in silicon production 2

Identifying structural defects in silicon production

 

A leading producer of medical imaging silicon wanted to reduce the number of defective wafers reaching the assembly stage.

Our objectives were to reduce the volume of faulty silicon assembled into equipment and reaching end customers by predicting which regions were defective after manufacture. The model pipeline was deployed into the client’s production process.

Implementation of image processing algorithms
Implementation of image processing algorithms 1
Implementation of image processing algorithms 2

Implementation of image processing algorithms

 

A leading producer of computer vision systems was looking into the possibility of porting their image processing software on to a new heavily parallelized processor architecture. This could offer a more configurable alternative to their existing FPGA implementation, without compromising performance; ultimately this provides a more open, customisable solution for their end-customers.