PDC Sensor for Cross-Traffic - Time-to-Intersection Estimation and Probabilistic Collision Risk Assessment for Ultrasonic Side Detection
This technical article explores the time-to-intersection estimation and probabilistic collision risk assessment for cross-traffic alert using PDC sensors. It covers the geometric modeling of intersection, the handling of measurement uncertainties, the fusion with camera for object classification, and the decision-making logic for issuing warnings and interventions.
The time-to-intersection (TTI) estimation for cross-traffic is based on a geometric model that assumes the vehicle is at the origin of a coordinate system, with the reverse direction as the positive x-axis. The approaching vehicle is detected at a distance (d) and an angle (θ) relative to the vehicle's longitudinal axis, measured using the time-difference-of-arrival (TDOA) between the multiple corner sensors. The lateral position (y) and longitudinal position (x) of the approaching vehicle are computed as x = d cos θ, y = d sin θ. The vehicle's speed is estimated from the rate of change of distance, and the velocity components are vx = dx/dt, vy = dy/dt. The intersection point is where the approaching vehicle's path crosses the vehicle's path (y=0). The TTI is computed as TTI = y / vy, assuming constant velocity. The system also accounts for the vehicle's own speed and steering, which changes the intersection point over time, using a dynamic model updated with the vehicle's CAN bus data.
PDC Sensor
Probabilistic collision risk assessment handles the uncertainties in the ultrasonic measurements, which include noise in distance and bearing, as well as the limited angular resolution (typically ±5°). The system models the measurement errors as Gaussian with known covariance matrices. Using a Monte Carlo simulation or an extended Kalman filter, the system estimates the probability distribution of the approaching vehicle's position and velocity. The collision probability is then computed as the probability that the vehicle will enter a certain danger zone (e.g., within 1 meter of the vehicle's path) within a future time horizon (e.g., 2 seconds). The warning is issued when the collision probability exceeds a threshold (e.g., 0.5) and the TTI is below a limit. The system also uses the quality of the measurement (SNR) to adjust the covariance, so that low-confidence measurements lead to a wider distribution and a lower probability threshold to still trigger warnings.
Fusion with a rear-view camera enhances the cross-traffic alert by providing object classification and visual confirmation. The camera detects and classifies the approaching object (vehicle, pedestrian, cyclist) using deep learning models. The classification is used to adjust the warning strategy: for pedestrians, the warning is more urgent due to higher vulnerability; for vehicles, the warning is based on TTI. The camera also provides lateral position with higher accuracy than the ultrasonic sensors, improving the bearing estimation. The fusion is performed at the object level: the ultrasonic track is associated with the camera detection using a nearest-neighbor algorithm based on position and velocity. The fused track then has a more accurate state estimate and a higher confidence, leading to more reliable warnings.
The decision-making logic for issuing warnings and interventions is hierarchical. The system first determines if the threat is credible (collision probability > 0.3). Then, based on the TTI and the classification, it selects the appropriate warning level (visual, audible, haptic). If the driver does not react within a certain time (e.g., 0.5 s after a high-level warning), and the collision probability exceeds 0.7, the system triggers automatic braking. The braking deceleration is modulated based on the TTI: for TTI < 0.5 s, full braking (0.8 g); for 0.5-1.0 s, moderate braking (0.4 g). The system also allows driver override by sensing pedal input. The logic is calibrated through extensive testing in various traffic scenarios to minimize false interventions while ensuring safety.
The cross-traffic alert system's performance is continuously improved through data-driven approaches. The system logs the detection events and the subsequent driver actions (braking, accelerating, steering) to refine the threat assessment parameters. Machine learning models are trained on large datasets to predict collision risk more accurately, considering the driver's behavior and the environmental context. The ultrasonic sensors themselves are being upgraded with higher bandwidth and improved signal-to-noise ratios to detect approaching vehicles at greater distances (up to 10 meters) and with better angular resolution, enabling earlier warnings. The combination of ultrasonic, radar, and camera technologies makes cross-traffic alert a robust and essential feature for modern vehicles, especially in urban environments where low-speed intersections are common.
