Table 1.

Summary of the deep-learning algorithms using different radar signal representations.

Reference	Radar Signal Representation	Network Model	Task	Object Type	Dataset	Remarks/Limitation
A. Angelov et al., [38]	Micro-Doppler signatures	CNN and LSTM	Target classification	Car, people, and bicycle	Self-developed	Their dataset is small. Hence, a larger radar dataset is required to train the neural network.
A. Danzer et al., [45]	Radar pointclouds	PointNet [149] and Frustum PointNets [35]	Car detection	Cars	Self-developed	Their dataset is relatively small, with only one radar object per measurement cycle. Besides, it contains a few object classes.
O. Schumann et al., [50]	Radar pointclouds	CNN, RNN	Segmentation and classification of static objects	Car, building, curbstone, pole, vegetation, and other	Self-developed	Their approach needs to be evaluated using a large-scale radar dataset.
O. Schumann et al., [51]	Radar pointclouds.	PointNet++ [145]	Segmentation	Car, truck, pedestrian, pedestrian group, bike, and static object	Self-developed	They used the whole radar point clouds as input, and obtained probabilities for each radar reflection point, thus avoiding the clustering algorithm. No semantic instance segmentation was performed.
S. Chadwick et al., [54]	Radar image	CNN	Distant vehicle detection	Vehicles	Self-developed	They used a very trivial radar image generation that does not consider the sparsity of radar data.
O. Schumann et al., [117]	Radar target clusters	Random forest classifier and LSTM	Classification	Car, pedestrian, bike, truck, pedestrian group, and garbage	Self-developed	Only classes with many samples returned the highest accuracy.
M. Sheeny et al., [122]	Range profile	CNN	Object detection and recognition	Bike, trolley, mannequin, cone, traffic sign, stuffed dog	Self-developed	Their system captured only indoor objects, and they did not make use of the Doppler information.
K. Patel et al., [123]	Range-Azimuth	CNN, SVM, and KNN	Object classification	Construction barrier, motorbike, baby carriage, bicycle, garbage container, car, and stop sign	Self-developed	Their system works on the ROIs instead of the complete rang-azimuth maps. And also the first to the allowed classification of multiple objects with radar data in real scenes.
B. Major et al., [124]	Range-azimuth-Doppler tensor.	CNN	Object detection	Vehicles	Self-developed	They showed how to leveraged the radar signal velocity dimension to improve the detection results
A Palffy et al., [125]	Range-Azimuth and radar Point clouds	CNN	Road users detection	Pedestrians, cyclists, and cars	Self-developed	They are the first to utilized both low-level and target-level radar data to addressed moving road user detection.
D. Brodeski et al., [126]	Range-Doppler-Azimuth-Elevation	CNN	Target detection	2-Classes object, and non-object	Self-built (in the anechoic chamber)	Real-world data was not included.
Y. Kim and T. Moon [139]	Micro-Doppler signatures	CNN	Human detection and activity classification	Human, dog, horse, and car	Self-developed	Their system could only detect humans presence or absence in the radar signal since there is no range and angle dimensions.
S. Lee [144]	Bird-eye- view	3D object detection	Cars	Cars	Astyx HiRes [39]	Radar Doppler information was not incorporated into the network.