Lightweight Key Point Detection Algorithm Design And Hardware Implementation

The key point detection refers to the detection of points with rich geometric features,such as the endpoints of lines,the center of circles,the intersections of lines,and the corner points.With the key point information extracted in adjacent frames,the robots or drones can compute the relative pose between the two frames accordingly,and then estimate the position of itself in the environment.In recent years,with the successful application of artificial intelligence in image processing tasks such as target detection,object classification and image segmentation,more and more researches attempt to adopt convolutional neural networks to improve the performance of key point detection.However,the heavy computational load and mass data transmission of convolutional neural networks make it difficult for such algorithms to be deployed on edge computing platforms such as robots and drones,where exactly the key point detection mainly applied.This research addresses the above-mentioned problems in a hardware software codesign fashion by proposing the following innovations: The research firstly proposes a hardware friendly key point detection algorithm based on lightweight convolutional neural network,which combines good effectiveness as well as efficient computational load and storage consumption.Secondly,the research designs a dedicated key point detection hardware accelerator with a novel loose-coupled multi-pipelines architecture,making the proposed key point detection algorithm to be processed in real time on FPGA platform.Experiments show that the proposed hardware can detect key points from the VGA resolution input frames in real time at 94 frames per second.The parameter size of the proposed algorithm is 28609 Bytes,which is much more efficient than other designs.Besides,the design also achieves a good ATE of 0.255 m when evaluated with public dataset EUROC,which shows a good accuracy and effectiveness.This research contributes to the deployment of convolutional-neural-network-based key point detections on edge computing platforms,and hence has a broad application prospect in Autonomous Mobile Robot(AMR)and Unmanned Aerial Vehicle(UAV).