Embedded Object Detection with Custom LittleNet, FINN and Vitis AI DCNN Accelerators

Object detection is an essential component of many systems used, for example, in advanced driver assistance (ADAS) or video surveillance (AVSS). Currently, the highest accuracy achieved by solutions using deep convolutional neural networks (DCNN). Unfortunately, these come at cost a high computational complexity; hence, work on widely understood acceleration algorithms very important and timely. In this work, we compare three different DCNN hardware accelerator implementation methods: coarse-grained (a custom called LittleNet), fine-grained (FINN) sequential (Vitis AI). We evaluate approaches terms object accuracy, throughput energy usage VOT VTB datasets. also present limitations each methods considered. describe whole process DNNs implementation, including architecture design, training, quantisation implementation. used two DNN architectures to obtain higher lower consumption. The first was implemented SystemVerilog second with FINN tool from AMD Xilinx. Next, both were compared Vitis AI final implementations tested Avnet Ultra96-V2 development board Zynq UltraScale+ MPSoC ZCU3EG device. For architectures, 196 fps our 111 FINN. same 123.3 53.3 fps, respectively.