Fast Drivable Area Detection Method Based On Heterogeneous Computing

Driving area detection means that the automatic driving system analyzes the area image taken in front of the vehicle to determine whether the front is a driving area without obstacles such as vehicles and pedestrians.The detection of drivable area is the basic component of the environment perception module of automatic driving.At present,the detection of drivable area based on deep learning has become the mainstream method in environmental perception.Compared with traditional detection methods,the method based on deep learning greatly improves the segmentation accuracy,can solve image processing problems in more complex environments,and can greatly improve the robustness of the algorithm to the environment.However,the drivable area detection method based on deep learning has higher computational complexity and a large amount of computation,and faces challenges in meeting application scenarios with high real-time requirements such as autonomous driving.The current mainstream computing methods based on ASIC,GPU,and FPGA have their own shortcomings,and it is still an area worthy of in-depth study to explore a better computing solution suitable for this application scenario.Aiming at the problems of low development efficiency and difficult algorithm iteration in drivable area detection based on field programmable gate array FPGA customizable calculation,this paper studies the FPGA custom computing method based on Vitis AI and DPU in solving the problem of drivable area detection.availability.The main work of this paper is as follows:(1)A drivable area detection algorithm based on improved U-Net is proposed.Firstly,based on the semantic segmentation method of convolutional neural network,deep learning is applied to the detection of drivable areas in street view images.Based on the end-to-end deep learning semantic segmentation model U-Net,the optimization strategy of convolution layer improvement,residual attention mechanism improvement,and dendritic network sampling is adopted to improve the detection accuracy of the algorithm for drivable areas of street view images.The highest overall accuracy rate of 95.1% is achieved on the dataset,which is 2.36%higher than the drivable area detection method before improvement,which significantly reduces the confusion between the drivable area and the background.(2)The FPGA-based drivable area detection hardware accelerator is studied.First,by analyzing the parallelism of convolutional neural networks,using loop unrolling technology and pipeline technology,a convolution operation module,a depthwise separable convolution module,and an activation function calculation unit module are designed.Then,the data fixedpoint quantization technology and the above hardware modules are used.Finally,the overall hardware design of drivable area detection is realized.(3)A fast detection system of drivable area based on FPGA is designed and implemented.In order to quickly implement a low-latency and low-energy drivable area detection architecture that is more efficient than the original software,making it suitable for low-power Io T and mobile devices,this paper studies the optimal parameter settings of DPU based on Vitis AI,and analyzes its Impact on system implementation latency,energy consumption,and resource utilization.Experiments show that the accuracy rate of the model after hardware acceleration remains at 94.3%,and the processing frame rate of 42 FPS and the performance of 1.95FPS/W can be obtained on the heterogeneous hardware platform based on ZCU 102.Compared with the i9-9900 X CPU with a frequency of 3.5GHz on a general computing platform,the processing speed is 3.2 times that of it,and the energy efficiency is improved by 24 times.Compared to the GTX 2080 GPU,it is 0.52 times faster and 5.4 times more energy efficient.