Optical Camera Communications with Convolutional Neural Network for Vehicle-toVehicle Links

This paper describes a vehicle-to-vehicle (V2V) communication system, employing optical camera communications (OCC). The system comprises the light emitting diode (LED)-based taillights and a raspberry camera used as the transmitter (Tx) and the receiver (Rx), respectively. The sectorized taillights (i.e., Tx) are intensity modulated at different frequencies, and a convolutional neural network (CNN) at the Rx is used for scene analysis, the region of interest (RoI) selection, and symbol detection. Results show that, the system data rates are constrained by the camera frame rate and symbol duration. The link performance is dependent on the CNN training set and we show that, the use of CNN allows a robust implementation, able to provide response under multiple situations: taillight obstruction, variable link distances, and misaligned Tx-Rx. Furthermore, CNN enables multiple input multiple output (MIMO) signal detection without the need for dedicated training.