Embedded implementation of a random feature detecting network for real time classification of time-of-flight SPAD array recordings

Abstract

A real time program is implemented to classify different model airplanes imaged using a 32x32 SPAD array camera in time-of-flight mode. The algorithm uses random feature extractors in series with a linear classifier and is implemented on the NVIDIA Jetson TX2 platform, a power efficient embedded computing device. The algorithm is trained by calculating the classification matrix using a simple pseudoinverse operation on collected image data with known corresponding object labels. The implementation in this work uses a combination of serial and parallel processes and is optimized for classifying airplane models imaged by the SPAD and laser system. The performance of different numbers of convolutional filters is tested in real time. The classification accuracy reaches up to 98.7% and the execution time on the TX2 varies between 34.30 and 73.55 ms depending on the number of convolutional filters used. Furthermore, image acquisition and classification use 5.1 W of power on the TX2 board. Along with its small size and low weight, the TX2 platform can be exploited for high-speed operation in applications that require classification of aerial targets where the SPAD imaging system and embedded device are mounted on a UAS.