Radar Based Object Recognition with Convolutional Neural Network

In this work, we investigate an application of convolutional neural networks (CNNs) using radar signal for recognizing objects and human movements. We suggested 7 scenarios and collected 2000 samples of each one: nothing, a carton, a plastic box inside the carton, a metal plate inside the carton, a man sitting behind the carton, a standing man and a moving man. The numeric radar data is preprocessed to generate image. Then it is sent to a pre-trained CNN (Inception V3) for feature extraction. The model is trained for 30 epochs in a batch of 20 samples. We choose to line-tune the top 2 inception blocks by training the model once again. After all this training process, it attains validation accuracy of 99.2% and testing accuracy of 99.7%. The model carries through the classification of these 7 scenarios.