Cadence Tensilica Vision DSPs for Imaging, Computer Vision, and Neural Networks
The scalable Cadence® Tensilica DSP cores and their software development environments offer ease of use which significantly reduces the development time of high-performance, low-power ADAS designs for deep learning/neural network vision processing applications. Demos include a 360° surround view, on-device Artificial Intelligence (AI) for people detection, and the Xtensa Neural Network Compiler (XNNC) for Image Classification.
360 Surround View
The 360 surround view image processing demo uses the Tensilica® Vision P6 DSP to provide a synthetic bird’s eye view of images captured by four cameras on an automobile, stitched together to form a top-view image using lens distortion correction and perspective transformation algorithms to create a single, seamless image that is displayed in an ADAS automotive application, for example in a park assist application.
On-Device AI for People Detection
The on-device AI for people detection uses the Tensilica® Vision P6 DSP and a You Only Look Once (YOLO) algorithm based on a neural network trained for people detection. YOLO is ideal for applications requiring fast, power-efficient object detection (with localization) for multiple object categories and can be retrained for up to 20 object classes. It uses custom network optimization, including pruning and quantization.
On-Device AI for Image Classification
The on-device AI for image classification demo features the Xtensa Neural Network Compiler (XNNC). The XNNC takes the Caffe or TensorFlow trained Inception V3 and Mobilenet neural networks and its floating-point coefficients (weights) and converts them to highly optimized fixed-point neural network code for a Tensilica® AI DSP. The XNNC drastically reduces the time to convert a trained neural network code for embedded devices from several months to several days! and also preserves the detection quality achieved in floating point in fixed point while providing the benefit of low power and lower memory bandwidth of the fixed-point processing.
For more information visit: https://ip.cadence.com/applications/automotive/ipg-automotive