In the intelligent driving scenario with limited on-board computing power, Hongmeng Next uses lightweight model technology to crack the dual challenges of performance and safety.This article disassembles the core scenario optimization strategies and practical solutions to help developers achieve efficient deployment.
1. Lightweight strategy for core scenarios
(One) Analysis of two major urgent needs scenarios
- Land Line Detection: It is necessary to deal with extreme conditions such as blurred rainy days and low light at night. The model needs to complete accurate identification of complex road conditions within 100ms.
- obstruction identification: It is required to warn 100 meters in advance for high-speed moving objects (such as 120km/h vehicles), and avoid misjudgment of road markings as obstacles.
(II) Three major optimization strategies
| Strategic Direction | Core Means | Effect Examples |
|---|---|---|
| Hardware adaptation | Adopting MobileNetV3 architecture + NPU acceleration | Reducing computing volume by 60% |
| Safety priority | 0.15 low-proportion pruning + 8-bit quantization | Accuracy loss is controlled within 1% |
| Distributed Collaboration | Camera end preprocessing + edge node inference | Data transmission volume is reduced by 75% |
2. Practical optimization of key functions
(I) Lane line detection speed-up plan
// MindSpore Lite lightweight process
const model = mslite.load('original model');
const pruned = mslite.prune(model, 0.2); // 20% structured pruning
const quantized = mslite.quantize(pruned, 8); // 8-bit quantization
quantized.save('lightweight model');
After optimization, the model volume was reduced from 12MB to 5MB, and the inference speed in complex curve scenarios was increased by 40%, and the accuracy rate was maintained at 92%.
(II) Balance technique for obstacle recognition
- Data Enhancement: Simulate 100+ scenarios such as rain and fog weather, dynamic occlusion, and expand the training data by 3 times
- Structural Optimization: Introducing Swin Transformer lightweight module, reducing the parameter volume by 35%
- Mixed Accuracy: Key layer FP16 computing, overall memory usage reduction by 50%
3. System-level assurance and future direction
(I) Reliability triple protection
- Dynamic backup: When the main model is abnormal, switch the backup model within 0.5 seconds
- Adaptive adaptation: Automatically adjust image preprocessing strategy according to camera parameters
- Redundant Design: Dual NPU chip hot standby, no sense of failure switching
(II) Technology evolution trend
- Soft and Hard Collaboration: Customized sparse computing architecture for Ascend automotive chips
- Multimodal Fusion: Develop a unified lightweight model that supports image + point cloud
- Reinforcement Learning: 60% reduction in decision model through end-to-end training
Top comments (0)