Hardware Oriented Vision System of Logistics Robotics

摘要

A qualified logistics robot is required to locate and identify the target item properly. The vision system of the robot is the way it perceives the world which demands high precision and low latency. Using the state-of-art deep convolutional neural network model, we present a folder detector to locate and identify the file folder. We implement several classic CNN models in the Faster RCNN framework. The average precision of the ideal MobileNet model is up to 0.966 with the GPU inference time 59ms. Besides the model design, we present a hardware oriented layer adaptive quantization method. Using this method, we condense the model into low bitwidth fixed point arithmetic which is more efficient and hardware friendly than the GPU widely used 32 floating point arithmetic. We can condense the model into 8-bit fixed point arithmetic without the precision drop.