ADC to Radar Cube

Raw ADC data is a stream of complex samples. It is not a point cloud and not an image. The first task is to reshape it into a frame structure that FFT processing can use.

radar_fft_cube_progress_en.ipynb uses this path. The aligned Chinese version is radar_fft_cube_progress_zh.ipynb.

read_dca1000_complex_bin
-> reshape_tdm_mimo_frames
-> range_fft
-> doppler_fft
-> angle_fft
-> detect_points_from_angle_cube

The parallel implementation splits this into dca1000_reader.py, fft_layers.py, point_cloud.py, and parallel_pipeline.py.

Dimensions

Dimension	Meaning	Used by
sample	ADC samples inside one chirp	Range FFT
loop / chirp	repeated chirps inside one frame	Doppler FFT
TX / RX	antenna channels	virtual antenna array
frame	time sequence	behavior modeling

After Angle FFT, the cube is organized as:

[doppler_bin, angle_bin, range_bin]

TX/RX Ordering

Raw ADC files are usually stored as a time-ordered stream, not as a clean tensor. The processing code needs radar configuration to recover:

[num_loops_per_frame, num_tx, num_rx, num_adc_samples]

In TDM-MIMO, TX antennas transmit one by one, and all RX antennas receive each chirp. If TX/RX ordering is wrong, range energy may still appear, but angle estimation will be unreliable because Angle FFT depends on phase relationships across the antenna array.

flowchart TB
  S["DCA1000 raw ADC bin<br/>interleaved I/Q samples"] --> R["read_dca1000_complex_bin"]
  R --> M["reshape_tdm_mimo_frames<br/>[frame, loop, tx, rx, sample]"]
  M --> RF["Range FFT<br/>axis = sample"]
  RF --> DF["Doppler FFT<br/>axis = loop"]
  DF --> AF["Angle FFT<br/>axis = virtual antennas"]
  AF --> P["detect_points<br/>range, velocity, angle, power"]