Rulin Liu, Zhihua Long
Aiming at the problems of insufficient real-time performance of multi-target detection and poor tracking stability in complex scenes in urban intelligent transportation system, this paper proposes a multi-target real-time recognition and tracking algorithm based on deep learning. Firstly, a target detection model integrating CNN network structure and attention mechanism is constructed, and the detection accuracy of vehicles, pedestrians and other targets is significantly improved by designing a novel SRFPN (SORT-CNN Feature Pyramid Network) module that integrates CNN and Transformer features to enhance multi-scale representation, and by optimizing the loss function with an additional penalty term specifically designed to reduce identity switches during data association. Secondly, the Simple Online and Realtime Tracking (SORT) algorithm framework is built and combined to form the algorithm model, so as to lighten the Re-ID feature extraction module, and finally, the target occlusion and ID switching problems are solved by using the trajectory prediction and data association optimization strategy; Experimental results show that the average detection accuracy of this algorithm is 95%, and the tracking accuracy is 82.4%, ranking first in the comprehensive evaluation index, which has significant advantages over traditional methods. The research results provide an effective solution for real-time target analysis in urban traffic monitoring scenarios, and have practical reference value for the construction of intelligent traffic management system.