自动泊车多任务轻量化感知模型研究

Research on Automatic Parking Multi-task Lightweight Perception Model

随着深度学习模型的发展,越来越多的模型用于各个行业,包括自动驾驶行业,但同时也面临着轻量化感知模型以及产品落地的挑战。然而在自动驾驶泊车感知中,常常受到光照、阴影等环境变化的影响,对空车位检测和可行使区域的识别是很大的难题。本文的算法通过车身左侧、右侧、前侧、后侧等四个车载180度广角鱼眼摄像头实时获取视频流,先将采集到的图像利用卷积神经网络对车辆所处区域周边各车位的状态、车辆可行驶的区域进行计算推理,然后对模型推理结果进行解析与融合。利用一个轻量级网络解决自动泊车过程中进行车位的感知和路面可行驶区域的感知,实验结果表明,帧率可达到19FPS,模型推理帧率可达到29FPS,单车位角中心点定位世界坐标系下平均误差为2.65cm,空车位检测成功率90%以上,满足实际应用对实时性、准确性、鲁棒性的要求。

With the development of deep learning models,more and more models are used in various industries,including the autonomous driving industry,but at the same time,they also face the challenges of lightweight perception models and product implementation.However,in the perception of autonomous parking,it is often affected by environmental changes such as lighting and shadows,and it is a great problem to detect empty parking spaces and identify the driving area.The algorithm in this paper obtains the video stream in real time through four 180-degree wide-angle fisheye cameras on the left side,right side,front side,and rear side of the vehicle,and first uses the convolutional neural network to calculate and reason the status of each parking space around the vehicle area and the area where the vehicle can drive,and then parses and fuses the model reasoning results.Using a lightweight network to solve the perception of parking spaces and the perception of road drivable areas in the process of automatic parking,the experimental results show that the frame rate can reach 19FPS,the model reasoning frame rate can reach 29FPS,the average error of single parking corner center point positioning world coordinate system is 2.65cm,and the success rate of empty parking space detection is more than 90%,which meets the requirements of real-time,accuracy and robustness in practical applications.