In state-of-art radar classification tasks the raw time domain ADC data is transformed to frequency domain wherein the target is detected. Feature data, such as the micro-Doppler signature, is extracted for classification of e.g. a performed gesture or a person based on its gait. However, the transformation into frequency domain using short-time Fourier transform resulting in erroneous target detection due to superposition of target components thus leading to sub-optimal feature for subsequent classification task. Thus, in this paper we propose a target feature extraction approach that operates directly on 2D time domain radar data by using a complex-valued adaptive 2D sine filter. The proposed approach tracks the target's slow-time and fast-time frequencies by a regulation loop, which progressively adjusts the filter's center frequency to the target location. We demonstrate that the proposed approach extracts better features in both single- and multi-target scenarios, leading to improved classification accuracy in gait classification problems using radars. Furthermore, the proposed time domain feature extraction approach facilitates the use of a parametric neural network that works directly on time domain data.