Research On Recognition Of Vehicle Drivable Area Based On Lightweight Neural Network

As the intersection of energy change and information change,autonomous driving has seen rapid development in research technology and wide application prospects,but at the same time autonomous driving still has many problems to overcome.With the further development of technology,the situations that autonomous driving needs to consider and deal with will be broader and more complex,from structured highways to unstructured natural area,from well-lit days to rainy and snowy nights,regardless of whether the application is for civilians,business or military.It is very important to research on scene recognition and driving area classification for complex environments.Supported on the National Key R＆D Program of Autonomous Driving Electric Vehicle Hardware-in-the-Loop Test Environment Construction and Simulation Test Technology Research（No.2018YFB0105103）,this paper researches the drivable area recognition problem of autonomous driving vehicles in complex environments,i.e.,image segmentation and target recognition,and conducts simulation experiments and real-world vehicle verification of the image segmentation algorithms.The research mainly includes the following points:1.Establish recognition algorithm based on maximum entropy for the recognition of drivable areas in complex scenes,considering the algorithm stability and computation consumption.Through noise reduction and cleaning of the input data,image pre-processing is completed to strengthen the information features.Introduce the concept of neighborhood pixel group grayscale,complete the calculation derivation of two-dimensional maximum Tsallis entropy based on the traditional image segmentation information entropy theory,propose an improved filtering template,and use the improved two-dimensional maximum Tsallis entropy algorithm to segment the input data for image recognition.2.Establish a lightweight convolutional neural network model for the recognition of drivable areas in complex scenes.The network structure is designed based on neural network and deep learning theory,the convolutional layer is designed based on residual module and channel-by-channel convolution,the parallel pooling module is designed based on the idea of parallel threads and processes in computer system and the spatial pyramid pooling theory,the improved activation function is proposed.Considering the relationship between algorithm accuracy,computation consumption and lightweight,adjust the network model through training and realize the recognition of drivable areas.3.Research and analyze the data distribution of current autonomous driving datasets and drivable areas datasets,and provide solutions to the problem of insufficient computer vision datasets in the face of subsequent research.Break up and fuse the public dataset to create a dataset of drivable areas recognition in complex situations.Collect data in the field and make a small dataset,expand and enhance the content of self-built datasets,consider the dataset as reference dataset for testing algorithms and adjusting neural networks.4.Propose an automated processing algorithm module that can be applied to process the current public dataset,and solve the situation that the data structure is not common and the labeling method is not uniform among the existing public datasets.The automated module processes relevant datasets and data information by acquiring the data structure of the dataset,cleaning and filtering data,and quickly modifying the label type,which effectively reduces the comprehensive labor consumption and time consumption when facing multiple datasets or large datasets,simplifies the process and improves work efficiency.5.Use the dataset and experimental platform to design relevant experimental scenarios,conduct simulations and real vehicle experiments,and verify the effectiveness and stability of the algorithm models and technical solutions.Two sets of image segmentation solutions are evaluated in the same scenario comparison experiments,the neural network model is verified according to the self-picked dataset,and the feasibility and effectiveness of the overall technical solution is tested and evaluated in real vehicles.This paper takes over part of the traditional image segmentation scheme theory,proposes a new solution,realizes dynamic segmentation recognition of drivable areas based on deep learning neural network in both virtual environment and reality,proposes editing optimization method for the shortage of public datasets,establishes selfcollection dataset,and completes the construction of drivable areas dataset,has various reference values at the practical application level.