Konferenzen und Rahmenprogramm
Hardware-Aware Grid Fusion Networks for Automotive
Deep convolutional neural networks (CNNs) offer a good alternative for classical sensor fusion algorithms, when relying on pre-processed sensor data. Recent research shows promising results, where feature-level inputs are used to generate an environmental model. To be applicable in real-world, the models have to be implemented on automotive ECUs with strict constraints. Therefore, existing approaches are analyzed in terms of computational burden and modified with modules known from SqueezeNet. Finally, the trade-off between parameter and MAC count reduction versus the degradation of fusion quality is evaluated. Fusion quality is measured with metrics known from pixel-wise classification tasks, like IoU and pixel accuracy. This paper shows that proposed grid fusion network preserves same level of fusion quality, while being compressed by factor of 5.7 in parameter size and 4.5 in MAC count. This compression allows proposed networks to run on Synopsys EV61 CNN engines in real-time.
--- Datum: 25.02.2020 Uhrzeit: 17:30 - 18:00 Uhr Ort: Conference Counter NCC Ost