Multi-vehicle detection in dynamic scenes: a Mask Region-Convolutional Neural Networks based study

In the context of complex dynamic scenes, continuous changes in scenes, mutual occlusion between vehicles, and other factors will cause many problems, such as low detection efficiency, false detection, and serious missed detection, so a vehicle learning algorithm based on deep learning is proposed. This algorithm is improved based on Mask_RCNN open source framework. Traffic scene image data in Dalian was obtained through cooperating with Dalian Traffic Police Detachment. And the data was labeled by referring to the labeling format of the COCO public data set. Training and test data set was constructed and the Mask_RCNN network model was improved. At last, taking advantage of the trained vehicle detection model, the detection of moving vehicles in complex scenes was realized. Experimental results show that with the help of the improved algorithm, the average detection speed increased from 4.12fps to 5.12fps on the basis of having an accuracy rate of 94.31% and AP of 74.7.