4 Deep Learning Acceleration

Next-generation CMS systems need to be able to identify vehicles, bicycles, and pedestrians reliably and have the capability of providing proximity warnings. Deep learning is highly effective for these tasks in the automotive context due to the capacity to handle variability such as scale, viewpoint, and lighting conditions thus allowing for robust detection performance. TI’s deep learning accelerator is the C7x, MMA DSP engine that is capable of 2 TOPs of performance. TI provides a model analyzer and model selection tool⁽²⁾ that enables third party perception stack providers to choose the deep learning model that provides the maximum entitlement in terms of frames per second and accuracy. As an example, Table 4-1 illustrates the performance entitlement with the SSDLite-MobDet-EdgeTPU model when running at 60 fps. This model is found in TI's edgeai-modelzoo⁽³⁾.