Abstract: Strong non-uniform, non-stationary and non-Gaussian clutter typically leads to low detection rate or high false-alarm rate for conventional threshold detection methods in energy domain. Target detection based on machine learning with the time-frequency features has been explored by some previous works. However, such features may be unavailable for short dwell-time in real searching applications, which results in the performance degradation or even failure for these methods. Therefore, a new radar target detection method with multi-frame multi-dimensional features based on a new deep neural network RDF-ResNet is proposed. By exploiting multi-frame multi-dimensional features of targets during ambiguity resolution, the proposed RDF-ResNet can suppress false alarms effectively. Together with low-threshold detection, a significant improvement of detection performance of radars can be achieved by the proposed detection method. Field experiment demonstrates that the proposed method can increase the detection rate by 41?? and reduce the false-alarm rate by 48% compared to the traditional threshold detection. A new way is also provided to establish feature spaces for radar echoes and detect targets with machine learning more effectively.