Lightweight Convolutional Neural Network Indoor Object Detection Algotithm Design And Hardware Acceleration

The total number of disabled people in China is more than 80 million,among which physical and visual disabilities account for the largest proportion.In order to solve the problem of difficulty in picking up the necessities of life for the visually and physically disabled,the latest research transplants the object detection algorithm to a mobile robot,which can effectively assist the visually and physically disabled in their daily activities by taking indoor photos with a camera and using the object detection algorithm to search for the location of the objects,solving the common problems of this type of disabled in their daily lives.In this paper,we investigate object detection models and hardware acceleration applicable to small indoor object datasets.In this paper,two lightweight object detection models,YOLOv3-ShuffleNetv2 and tiny-YOLOv3-ShuffleNetv2,are proposed based on the YOLOv3 model.To verify the effectiveness of the models,an indoor small object dataset,indoor2020,is created and used as a criterion to evaluate the performance of the models,comparing detection detection accuracy and speed of YOLO v3,YOLOv3-ShuffleNetv2 and tiny-YOLOv3-ShuffleN-etv2.To further improve the detection speed,the SoC hardware design of each module of the object detection model is accelerated using the HLS tool.The paper uses a scheme that combines ShuffleNetv2,a lightweight network,with YOLOv3 which is a object detection network.To ensure better network detection and to lighten the model so that it meets the lightweight requirements for porting to mobile robots,and to improve and optimize the model for data images so that it can be applied to small indoor object datasets.The hardware acceleration part uses the Zynq-7000 series SoC demo board to design the trained model in modules.The object detection model designed in this paper mainly contains standard convolution,deep separable convolution and maximum pooling layer,so the hardware acceleration part carries out hardware structure building and simulation synthesis for these three modules respectively.The experimental results show that the YOLOv3 model size is 235M and the detection speed is 12 fps.The YOLOv3-ShuffleNetv2 model combining YOLOv3 and ShuffleNetv2 has a size of 82M and a detection speed of 26 fps.Compared with YOLOv3,the model is reduced to 1/3 and the speed is doubled.The tiny-YOLOv3-ShuffleNetv2 model for small target datasets has a size of only 20M and a detection speed of up to 38 fps.Compared with YOLOv3,the model is reduced to 1/10 and the speed is increased by 3 times.Evaluated on the indoor dataset,the YOLOv3 model mAP was 0.61,the YOLOv3-ShuffleNetv2 model mAP is 0.65 and the tiny-YOLOv3-ShuffleNetv2 model mAP is 0.78.Simulations and synthesis show that all three convolutional cores have a small resource utilization of no more than 10%on the SoC demo board.It can be seen that the tiny-YOLOv3-ShuffleNetv2 model proposed in this paper can meet the requirements of real-time and lightweight,and the accuracy rate is the highest among the three models,which can achieve the goal of transplanting into mobile robots.The object detection model and hardware acceleration scheme designed in this paper can provide algorithm and hardware implementation reference for the detection of small indoor object datasets.