DEV Community

lyc233333
lyc233333

Posted on

Hongmeng Next Intelligent Driving: A Practical Guide to Lightweight Models

In the scenario where on-board computing power is limited, Hongmeng Next uses model lightweight technology to crack the performance and safety challenges of intelligent driving.This article disassembles the core scenario optimization strategy, attaches practical solutions and performance data to help developers achieve efficient deployment.

1. Core scenarios and lightweight strategies

(One) Two key scenarios

  1. Land Line Detection: Under extreme conditions such as heavy rain and night, lane line identification must be completed within 100ms to deal with interference such as reflection and occlusion.
  2. obstruction identification: Warn 100 meters in advance for high-speed moving objects at 120km/h to avoid misjudging road signs as obstacles.

(2) Three-dimensional optimization strategy

Strategic Direction Core Means Typical Effects
Hardware adaptation MobileNetV3 architecture + NPU acceleration Computing volume is reduced by 60%
Safety priority 15% low proportion pruning + 8-bit quantization Accuracy loss ≤1%
Distributed Collaboration Camera end preprocessing + edge node inference Data transmission volume is reduced by 75%

2. Practical combat for key technologies optimization

(I) Lightweight lane line detection 

// MindSpore Lite optimization process
const model = mslite.load('original model');
const pruned = mslite.prune(model, 0.2); // 20%结构化剪枝  
const quantized = mslite.quantize(pruned, 8); // 8-bit quantization
quantized.save('lightweight model');
Enter fullscreen mode Exit fullscreen mode

Optimization results: The model volume has been reduced from 10MB to 6MB, the inference speed has been increased by 30%, and the accuracy of complex road conditions has been maintained at 90%+.

(II) Obstacle identification and balance technique

  • Data Enhancement: Simulate 100+ scenarios such as rain and fog, dynamic occlusion, and expand the training data by 3 times
  • Structural Optimization: Introduce a lightweight attention module, and reduce the amount of parameters by 35%.
  • Mixed Accuracy: Key layer FP16 computing, memory usage reduction by 50% Performance: 95% recognition accuracy +, 40% increase in inference speed, meeting real-time requirements.

3. System integration and reliability guarantee

(I) Key points of integration

  1. 分布式部署：车道线模型部署于摄像头节点，障碍物模型部署于车载NPU，通过鸿蒙分布式通信协同
  2. Adaptive adaptation: Automatically adjust image preprocessing strategy according to camera parameters to solve the adaptation problems of different models

(II) Reliability triple protection

  1. Dynamic backup: Switch the backup model in 0.5 seconds when the main model is abnormal
  2. Hardware redundancy: Dual camera hot backup, single sensor failure does not affect system operation
  3. Robustness Optimization: Adversarial training enhances extreme lighting adaptability, and improves accuracy to 90%

4. Future technology trends

  1. Soft and Hard Cooperation: Customize the sparse computing architecture for Ascend's automotive chips to further improve efficiency
  2. Multimodal fusion: Develop a unified lightweight model that supports image + point cloud to reduce the complexity of fusion
  3. Reinforcement learning integration: End-to-end training achieves 60% reduction in decision-making model, improving intelligent decision-making capabilities

Top comments (0)