Non-Motorized Lane Target Behavior Classification Based on Millimeter Wave Radar With P-Mrca Convolutional Neural Network

In the fields of road regulation and road safety, the classification of target behaviors for non-motorized lanes is of great significance. However, due to the influence of adverse weather and lighting conditions on the recognition efficiency, we use radar to perform target recognition on non-motorized lanes to cope with the challenges caused by frequent traffic accidents on non-motorized lanes. In this paper, a classification and recognition method for non-motorized lane target behavior is proposed. Firstly, a radar data acquisition system is constructed to extract the micro-Doppler features of the target. Then, in view of the shortcomings of traditional deep learning networks, this paper proposes a multi-scale residual channel attention mechanism that can better perform multi-scale feature extraction and adds it to the convolutional neural network (CNN) model to construct a multi-scale residual channel attention network (MrcaNet), which can identify and classify target behaviors specific to non-motorized lanes. In order to better combine the feature information contained in the high-level features and the low-level features, MrcaNet was combined with the feature pyramid structure, and a more efficient network model feature pyramid-multi-scale residual channel attention network (P-MrcaNet) was designed. The results show that the model has the best scores on classification indexes such as accuracy, precision, recall rate, F1 value and Kappa coefficient, which are about 10% higher than traditional deep learning methods. The classification effect of this method not only performs well on this paper’s dataset, but also has good adaptability on public datasets.