| Millimeter wave radar is a high-precision sensor used to measure the relative distance,velocity and orientation of objects.It has many advantages,including highprecision measurement results,strong anti-interference ability,all-day and all-weather availability,high-resolution detection of multiple targets,excellent performance in sensitivity and false alarm rate,high frequency and low power,simultaneous measurement of speed and distance,and long measurement range and high real-time performance.These advantages make millimeter-wave radar have great application value in the field of target detection.The existing millimeter-wave radar target detection algorithm usually first performs Fourier transform on the data,then performs Constant False-Alarm Rate(CFAR)detection,and then performs feature extraction and target detection on the detected point cloud data.This may lead to the loss of some effective information in the millimeter wave radar signal,resulting in missed detection.With the development of deep learning technology in the past decade,the original millimeterwave radar data after Fourier transform is directly input into the neural network,skipping CFAR processing to learn the characteristics of the target in the visual image,and the end-to-end target detection method has been widely studied.RODNet(RealTime Radar Object-Detection Network)is one of the representatives.Its advantage is that it directly inputs data into the deep network and retains effective information to the maximum extent,but there is a problem of missed detection of dense target scenes.The labeled radar data sets such as CRUW,KITTI and KAIST are also published.RODNet is one of the most widely used radar data target detection algorithms.The algorithm is based on Convolutional Neural Networks(CNNs),and uses millimeter-wave radar Range-Azimuth(RA)image sequence as input to directly learn radar RA image sequence data features,classify and locate targets in the scene,and complete target detection tasks.However,RODNet uses an OLS(Object Location Similarity)detection method that is independent of the number of targets.This method obtains the final target detection result from the predicted Confidence Map.Therefore,the missed detection rate of the algorithm is high in the dense pedestrian scene.In this paper,the problem of missed detection of RODNet in dense pedestrian scenes is analyzed and studied.The specific research work is as follows :(1)This paper analyzes the spatial distribution characteristics of Conf Map output by RODNet and the performance of OLS-based location processing methods in various scenarios.The limitations of OLS-based methods are summarized for adjacent pedestrian scenarios,and the main factors causing missed detection of adjacent pedestrian scenarios are clarified.(2)Aiming at the problem that the OLS method does not consider the spatial distribution characteristics of Conf Map and only considers its numerical distribution,this paper proposes a Gaussian Mixture Model with Target Number(GMM-TN).This method traverses different target number assumptions,clusters the Conf Map data generated by the RODNet network to obtain the target center,and simulates the Conf Map corresponding to the number of targets in the hypothesis space.(3)The KL(Kullback-Leibler)divergence is introduced as the criterion of spatial distribution similarity,and the RODNet millimeter wave radar target detection method based on GMM-TN is constructed.Through the CRUW data set,the target detection experiment is carried out on the adjacent pedestrian scene,and the confidence distribution under the typical hidden variable condition is analyzed.By comparing the detection results of one-dimensional and two-dimensional KL divergence as the target number estimation criterion with the detection results of RODNet,the effectiveness of the improved algorithm is verified : the two indicators of one-dimensional KL divergence are the highest,the average precision(AP)is increased by 29 %,and the average recall rate(AR)is increased by 36 %. |
PDF Full Text Request