Automatic Traffic State Recognition from Videos Based on Autoencoder and k-Means Clustering

An automatic video traffic state recognition method based on an autoencoder and k-means clustering is proposed to timely and effectively recognize road traffic state. First, candidate autoencoders are established through a reasonable optimization of the structural parameters on the input data dimensions, number of hidden layers, and dimensions of dimension-reduced data through cross-examination. Then, three image data sets are formed with 1500-4500 sample images. On this basis, the candidate autoencoders are trained and tested; thus, the best autoencoder AE^* is proposed according to precision, recall, and F1-value. Lastly, four traffic state recognition models are constructed by combining AE^* with k-means clustering, Support Vector Machine (SVM), Linear Classifier, and DNN Linear Classifier (Deep Neural Network with Linear Classifier), which are named as AE^*-kmeans, AE^*-SVM, AE^*-Linear, and AE^*-DNN_Linear, respectively. The models are trained and tested on the basis of the three image data sets. Results show that the four models' average precision in terms of precision and recall is 91.9%-92.7%, and their average recall is 91.6%-92.6%, while AE^*-kmeans performs best or second to best in terms of precision and recall. With regard to the comprehensive evaluation index F1-value, AE^*-kmeans achieves 92.4%, a little lower than 92.7% of AE^*-SVM, and better than AE^*-DNN_Linear (92.1%) and AE^*-Linear (91.8%). Given that k-means is an unsupervised clustering method, compared with AE^*-SVM, AE^*-Linear, and AE^*-DNN_Linear, AE^*-kmeans can reduce the workload, such as manual data calibration and supervised training and cut down calculation cost. Meanwhile, AE^*-kmeans also obtains a good traffic state recognition result. Therefore, this mechanism has high practical significance for an accurate real-time extraction of a video traffic status.