How to deploy the deep learning model to the actual project? (classification + detection + segmentation)

Application background

As early as far away 1989 year , One is called ALVIVN For the first time, the company has used neural networks in automobiles , Carry out lane line detection and ground segmentation . today , Deep learning has been applied in many branches of auto drive system . The first is the field of perception , Commonly used sensors are cameras 、 Lidar and millimeter wave radar . Deep learning uses two-dimensional images or three-dimensional point clouds as input , Detect the obstacles 、 distinguish 、 Division 、 Tracking and ranging .

Mask RCNN

M3DSSD: Monocular 3D Single Stage Object Detector

PointPillars: Fast Encoders for Object Detection from Point Clouds

The second is the positioning field , Automatic driving usually requires centimeter level positioning accuracy , This makes the traditional high-precision map unreliable in many scenarios . In recent years, some methods use online map learning , Based on on-board sensor observation , Build high-definition maps dynamically , It is a more scalable way than traditional pre annotated HD maps , Provide semantic and geometric priors for autonomous vehicles .

HDMapNet: An Online HD Map Construction and Evaluation Framework

Thirdly, forecast planning , Using deep learning method can better predict the trajectory of obstacles , There are even ways to put perception - Combine forecasting to solve problems .

PnPNet: End-to-End Perception and Prediction with Tracking in the Loop

Because deep learning algorithm is widely used in automatic driving , This makes model deployment engineers hot , Most companies require Algorithm Engineers to design algorithms , Deploy migration again , Talents with both skills have always been the preferred candidates for automatic driving companies , The corresponding salary is also very considerable .