Multitarget Detection in Depth-Perception Traffic Scenarios

Abstract

Multitarget detection in complex traffic scenarios usually has many problems: missed detection of targets, difficult detection of small targets, etc. In order to solve these problems, this paper proposes a two-step detection model of depth-perception traffic scenarios to improve detection accuracy, mainly for three categories of frequently occurring targets: vehicles, person, and traffic signs. The first step is to use the optimized convolutional neural network (CNN) model to identify the existence of small targets, positioning them with candidate box. The second step is to obtain classification, location, and pixel-level segmentation of multitarget by using mask R-CNN based on the results of the first step. Without significantly reducing the detection speed, the two-step detection model can effectively improve the detection accuracy of complex traffic scenes containing multiple targets, especially small targets. In the actual testing dataset, compared with mask R-CNN, the mean average detection accuracy of multiple targets increased by 4.01% and the average precision of small targets has increased by 5.8%.