radar

**Radar (Radio Detection And Ranging)** uses transmitted electromagnetic waves and their echoes to detect, locate, and characterize targets. Modern automotive and military radar systems are among the most demanding signal-processing applications, requiring real-time processing of 3D point clouds at millimeter precision. **FMCW (Frequency Modulated Continuous Wave)** is the dominant waveform for short-to-medium range radar. A linear chirp sweeps bandwidth B over time T; mixing the echo with the transmitted signal produces a beat frequency fb proportional to range (R = c·fb·T/2B). A 2D FFT over fast-time (range) and slow-time (Doppler) dimensions produces the Range-Doppler map showing both position and velocity of targets simultaneously. **Phased arrays** steer beams electronically by applying differential phase shifts across antenna elements, achieving millisecond beam switching versus mechanical seconds. MIMO radar multiplies virtual aperture: Tx_count × Rx_count virtual elements dramatically improve angular resolution without additional hardware. **CFAR (Constant False Alarm Rate)** detection adaptively thresholds the range-Doppler map by estimating local noise from surrounding cells. CA-CFAR averages reference cells; OS-CFAR (ordered statistic) is more robust to clutter edges. The goal is constant false alarm probability regardless of varying noise floor. **Key waveform parameters**: Range resolution ΔR = c/2B (finer with wider bandwidth); velocity resolution Δv = λ/2NT; maximum unambiguous range Rmax = c·T/2; maximum unambiguous velocity vmax = λ/4T. These create fundamental trade-offs: wider bandwidth → better range resolution but more ADC bandwidth; longer coherent integration → better velocity resolution but slower update rate. **AI/ML context**: Deep learning is transforming radar signal processing at every layer. CNNs classify targets from micro-Doppler signatures (pedestrian gait, hand gestures). PointNet architectures process sparse radar point clouds. Transformer-based sensor fusion combines radar, LiDAR, and camera for L4 autonomous driving. Radar SoCs (TI AWR, NXP S32R) now integrate ARM cores with hardware accelerators for on-chip neural network inference. ```svg FMCW Waveform + Range TX Chirp (Frequency vs Time): t f TX RX (delayed τ) fb T = chirp period τ = 2R/c (round-trip delay) fb = (B/T)·τ (beat freq) R = (c·fb·T)/(2B) B = bandwidth; c = 3×10⁸ m/s Range Resolution ΔR = c / (2B) 77 GHz, B=4 GHz → ΔR = 3.75 cm 24 GHz, B=250 MHz → ΔR = 60 cm FMCW Signal Chain: TX Mix ADC FFT→CFAR ↑ RX Range-FFT → Doppler-FFT → CFAR detect Range-Doppler Map Doppler (velocity) → Range → ground clutter v=0 -vmax +vmax R_max R_min Car (0m/s) Oncoming Ped. fd = 2·v·fc/c (Doppler shift) Δv = λ/(2·N·T) (velocity res.) Phased Array + MIMO TX Phased Array (4-element): PA φ1 ant PA φ2 ant PA φ3 ant PA φ4 ant Phase taper → steers beam angle θ θ = arcsin(Δφ·λ/(2π·d)) MIMO Virtual Aperture: Tx antennas × Rx antennas = virtual elements (angle res. ↑) 4Tx × 4Rx = 16 virtual elements Angular res. θ_res = λ/(N_virtual·d) Radar Frequency Bands: 24 GHz ISM: parking sensors, SRR 77 GHz auto: ACC, AEB, L3/L4 ADAS 94 GHz W-band: high-res imaging 300 GHz D-band: sub-mm research AESA defense: L/S/X/Ku bands CFAR Detection CA-CFAR (Cell-Averaging): Estimate noise from guard + reference cells Threshold = α × mean(reference cells) Detect if CUT > threshold (const. FA rate) OS-CFAR (Ordered Statistic): Sort cells, take k-th order stat as threshold Robust to clutter edges + interferers AI-CFAR (Deep learning): CNN replaces hand-crafted threshold rule Learns clutter statistics from training data 3dB lower detection threshold at same PFA Radar Equation + Parameters Radar Range Equation: Pr = Pt·Gt·Gr·λ²·σ / (4π)³·R⁴·L Pt: transmit power (dBm) G: antenna gain (dBi) σ: target RCS (m²) R: range (m) — note R⁴ dependence! L: system losses (dB) Key metrics: SNR = Pr / (kTBF) — sets Pd/PFA RCS of car: ~10 m², pedestrian: ~0.1 m² ROC curve: Pd vs PFA tradeoff AI-Augmented Radar Object classification: PointNet on radar point cloud: car/ped/cyclist Micro-Doppler CNN: breathing, gait, gesture Sensor fusion (ADAS): Radar + LiDAR + Camera → transformer fusion Radar works in rain/fog where camera fails Radar SoC (AI chip): TI AWR2944: ARM R5 + DSP + HWA on 28nm NXP S32R41: 16nm FinFET, CNN accelerator Hailo + radar: edge inference at 26 TOPS Imec 140GHz CMOS: 4D radar on single chip ```

Go deeper with CFSGPT

Get AI-powered deep-dives, save terms, and run advanced simulations — free account.

Create Free Account