Based on the motion state detection and obstacle detection neural network algorithm, we capture the computer game screen (obstacle scene) in real-time through the CIS camera and perform preprocessing on the FPGA. Then, we use the low-power burst neural network architecture to calculate the obstacle avoidance information and finally display the obstacle avoidance results with real-time game screens.
The technique can be applied to reversing warning systems and drone landing systems.

1. Low-power intelligent image sensing system:Propose a time-contrast pixel architecture and an exposure compensation mechanism, which is currently the simplest architecture for inter-pixel intra-frame difference calculations.2. Obstacle detection and obstacle avoidance neural network:A complete obstacle avoidance solution is compatible with vision, computing platforms, and unmanned vehicle hardware. The number of obstacle avoidance network nerves is less than 500.

The advantages of the team＇s target development module include fast speed, low power consumption, high integration, small size, and a single module that has the function of avoiding obstacles. To accurately display the above advantages in a limited space, using computer game screens to display real-time obstacle avoidance results will be further applied to smart manufacturing, smart agriculture, and smart home: for example, patrolling orchard farmland, flexibly driving away birds and macaques.