How Mobile AI Transforms Logistics: Real‑World Image Algorithms at Huolala

Introduction

Artificial intelligence (AI) has permeated every industry, and mobile AI is emerging as a powerful capability for devices. Unlike traditional cloud‑based AI that requires heavy compute and fast networks, mobile AI runs directly on sensors, cameras, and smartphones, offering real‑time processing, privacy protection, and efficiency gains.

Application Cases

Document Verification

Driver onboarding at Huolala requires document verification. Previously, photos were uploaded to the cloud, and poor‑quality images caused rejections and delays. Mobile AI now provides instant quality feedback during capture, improving photo acceptance rates and speeding up driver onboarding.

Vehicle Sticker Inspection

To ensure correct placement of brand stickers, Huolala conducts random inspections. Drivers must upload photos from specific angles. Mobile AI guides drivers in real time to capture compliant images, reducing rejection rates and associated costs.

Algorithm Solutions

Huolala adopts innovative approaches for mobile AI image algorithms, including deep model design, hybrid deep learning and traditional image processing, model lightweighting, data‑flow handling, and mobile deployment.

Deep Model Design

For sticker inspection, object detection is used. Popular methods such as YOLO, SSD, and DETR are referenced. Huolala extends SSD with depthwise separable convolutions and a lightweight detection head to also predict vehicle angle, adding angle labels during training.

Combining Deep Learning with Traditional Image Processing

In document verification, deep models first detect whether the uploaded document meets requirements, then traditional image‑processing algorithms assess blur or glare, providing immediate feedback to drivers for corrective actions.

Model Lightweighting

To meet diverse vehicle types, lighting, and hardware constraints, Huolala applies knowledge distillation to compress models. The student model runs at FP32 precision, achieving sub‑30 ms latency on Snapdragon 865 with a size of only 357 KB.

Recent advances in Vision Transformers (ViT, MobileViT, TinyViT) enable real‑time inference on mobile CPUs and GPUs, further expanding mobile AI capabilities.

Data Stream Processing

Huolala aligns per‑frame inference results with image streams, aggregating multiple frames to provide real‑time feedback during sticker capture.

Mobile Deployment

Given heterogeneous mobile hardware, Huolala supports multi‑backend inference (CPU/GPU) and provides a cloud‑based model update interface for seamless upgrades.

Conclusion and Outlook

The paper outlines Huolala's practical exploration of mobile AI image algorithms across document verification and sticker inspection, covering model design, compression, deployment, and data handling. Future work includes extending to 3D vision and AR to further enhance logistics matching and safety.

References

Redmon J, Farhadi A. Yolov3: An incremental improvement. arXiv preprint arXiv:1804.02767, 2018.

Liu W, Anguelov D, Erhan D, et al. SSD: Single shot multibox detector. Computer Vision–ECCV 2016, 2016: 21‑37.

Carion N, Massa F, Synnaeve G, et al. End‑to‑end object detection with transformers. European conference on computer vision, 2020: 213‑229.

Howard AG, Zhu M, Chen B, et al. MobileNets: Efficient convolutional neural networks for mobile vision applications. arXiv preprint arXiv:1704.04861, 2017.

Hinton G, Vinyals O, Dean J. Distilling the knowledge in a neural network. arXiv preprint arXiv:1503.02531, 2015.

Vaswani A, Shazeer N, Parmar N, et al. Attention is all you need. Advances in neural information processing systems, 2017, 30.

Dosovitskiy A, Beyer L, Kolesnikov A, et al. An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929, 2020.

Mehta S, Rastegari M. MobileViT: Light‑weight, general‑purpose, and mobile‑friendly vision transformer. arXiv preprint arXiv:2110.02178, 2021.

Wu K, Zhang J, Peng H, et al. TinyViT: Fast pretraining distillation for small vision transformers. European Conference on Computer Vision, 2022: 68‑85.