Yolo-based Realtime Object Detection using Interleaved Redirection of Time-Multiplexed Streamline of Vision Snapshot for Lightweighted Embedded Processors

The most essential function for fully automated driving is object recognition, as the car has to avoid obstacles. However currently, applying the technology can be challenging as they cost a lot of time and has lack of accuracy for embedded processors for automobiles. In this paper, we suggest time-multiplexed streamline of vision snapshot based on YOLOv3 for lightweighted embedded processors. As YOLOv3 process image detection with a single neural network in 1 evaluation, it can run fast. However, for lightweighted embedded processors, YOLO running on real time still can be heavy. By training YOLOv3 and Tiny YOLOv3, we lightened image detection program. As there are some latancies while running YOLO only on the lightweighted processor, ls1028 , we designed a socket communication program to communicate with tiny edge device based on python. The tiny edge device is connected with webcam and it sends snapshot to the processor by socket. When processor receives the snapshot, it detects objects using tiny YOLOv3. Then processor sends the coordinate information to tiny edge device using socket. Tiny edge device can draw a rectangle line above detected object. As a result, by using this system, we reduced webcam latency and some delay time drawing lines above the object and able to run YOLO on real time with lightweighted embedded processor.