Ultra-Low-Power ADAS on DiDi's JueShi Devices for Reducing Traffic Accidents

Road traffic accidents cause massive loss of life and property each year. As a deep participant in the transportation field, DiDi continuously explores ways to reduce these accidents. This article introduces how DiDi's vehicle‑vision team leverages the ultra‑low‑power JueShi ADAS (Advanced Driver‑Assistance System) on JueShi devices to lower accident rates and protect driver and passenger safety.

Background : Rear‑end collisions account for 60% of accidents with the same‑level responsibility, and 80% of those are caused by following too closely. Studies show that a 1.5‑second pre‑warning can prevent 90% of crashes, and forward‑collision warning (FCW) can reduce non‑braking rear‑ends by 69%.

DiDi’s JueShi ADAS integrates front and rear cameras, IMU, GPS, and runs multiple algorithms (DMS, collision detection, driver‑behavior detection) on edge devices to identify hazardous scenarios and intervene in real time.

Ultra‑low‑power front‑car detection : Traditional heavy models (e.g., TridentNet with ResNet‑101) are too slow for real‑time use on edge hardware. Lightweight backbones such as ShuffleNetV2 + SSD are adopted, and a custom anchor‑based regression model (ZoomNet) is designed to detect the preceding vehicle with minimal computation, occupying less than 5% of the CPU on an MTK8665 quad‑core ARM Cortex‑A53 processor.

During inference, the model predicts offsets for a set of anchors (every 120 px) and averages them to obtain the final bounding box, achieving real‑time performance.

Stability improvements : Lightweight models suffer from small‑object detection failures, large bounding‑box errors, missed detections, and instability under challenging conditions (night, rain, glare). To address this, a temporal‑fusion strategy combines two deep networks with a Kalman filter. The system first estimates a coarse position, refines it with a regression network, and predicts future positions to stabilize detections, reducing bounding‑box jitter by 23.3%.

Camera installation calibration : Varying mounting angles affect distance estimation. Two methods are proposed: (1) deep‑learning‑based vanishing‑point regression to calibrate pitch; (2) statistical estimation of the horizontal vanishing point from long‑term model outputs to infer yaw, enabling accurate longitudinal distance calculation.

Timeliness of alerts : Time‑to‑collision (TTC) is computed as distance divided by relative speed, with a typical threshold of 2.7 s. To improve TTC in high‑speed scenarios, a brake‑light classification model detects sudden braking of the preceding vehicle, triggering a front‑brake warning (BLW) and extending the effective warning window.

Long‑tail data mining : The platform collects collision, driver‑behavior, and DMS data at scale. By analyzing pre‑alert and post‑alert data, it identifies missed alerts, false positives, and hard samples where alerts did not prevent accidents. These samples feed back into model training and OTA updates, continuously improving coverage of complex scenarios.

Conclusion : The JueShi ADAS system demonstrates that with ultra‑low‑power edge AI, large‑scale deployment can significantly reduce rear‑end collisions (11.4% reduction in rear‑end rate, 9.1% overall accident reduction) and improve driver safety. Future work includes expanding model coverage, adding pedestrian collision warning (PCW), and further optimizing long‑tail performance.