ket
In recent times, automotive sensing application needs – especially for computer vision – have pushed the on-board compute demand significantly. The majority of today’s solutions have been designed and architected for non-automotive applications, and then repurposed for automotive use cases. These solutions create inefficient system architectures which result in ultra-high power and cost, and low system reliability. As a result, they are mainly used in premium vehicles without a path to scale them into mass market segments. As OEMs adopt increasingly higher resolution image sensors for CV and human vision, the processing requirement through the convolutional neural network for object detection grows exponentially. We will discuss a CV data path that uses a dedicated pre-processing imaging pipeline that is built for real time scene specific image signal processing with color management and significantly enhances the performance of the CNN engine by creating sharp contrast, scaled down, region-of-interest sub-images for efficient low-power processing. This allows the CNN to operate at reduced frame size and rate, thus reducing the overall processing and power requirements. As a result, the processor delivers the most power efficient performance per TOPS, thereby also enabling the use of the processor in constrained environments such as camera heads or a small fusion box.