基于迷你卷积神经网络的停车场空车位检测方法

针对日益严峻的停车难问题，提出一种基于改进卷积神经网络停车场空车位检测方法。首先，根据车位只需用两种状态来表示其占空的特点，对传统卷积神经网络结构进行改进，提出迷你卷积神经网络（MCNN）的概念；然后，通过减少网络参数来减少训练和识别时间，并在网络中加入局部响应归一化层以加强对明度的校正，以及使用小卷积核来获取更多图像细节；最后，对视频帧图进行手动掩码设置，通过边缘检测切割成单个车位图，并使用训练好的MCNN进行车位识别。实验结果表明，与传统机器学习方式相比，基于MCNN的检测方法识别率能提高3~8个百分点，同时网络参数仅为常规使用卷积模型的1/1 000，且在文中所述的几种不同环境中，识别率的均保持在92%以上。实验结果表明，MCNN可移植到低配置摄像头，实现停车场空车位自动检测。

Parking lot space detection method based on mini convolutional neural network

For the increasingly severe parking problem, a method of parking lot space detection based on a modified convolutional neural network was proposed. Firstly, based on the characteristic that a parking lot only needs to be denoted by two states, a concept of Mini Convolutional Neural Network (MCNN) was proposed by improving the traditional CNN. Secondly, the number of network parameters was decreased to reduce the training and recognition time, a local response normalization layer was added to the network to enhance brightness correction, and the small convolution kernel was utilized to get more details of the image. Finally, the video frame was manually masked and cut into separate parking lots by edge detection. Then the trained MCNN was used for parking lot recognition. Experimental results show that the proposed method can improve the recognition rate by 3-8 percentage points compared with the traditional machine learning methods, and the network parameters of MCNN is only 1/1000 of the conventionally used convolutional model. In several different environments discussed in this paper, the recognition rate maintains above 92%. The experimental result shows that the MCNN can be transplanted to a low-configuration camera to achieve automatic parking space detection.