In the field of smart transportation, traditional object detection and tracking methods face many difficulties in complex traffic environments, especially in effectively identifying and tracking partially occluded targets. This article intends to take Faster R-CNN (Faster Region-based Convolutional Neural Networks) as the research object to study multi-objective tracking algorithms for local occlusion, in order to improve the operational efficiency and safety of intelligent transportation systems. Firstly, this article intends to study a new deep learning based object detection method for solving a large number of occluded pedestrians and vehicles in complex traffic environments. Secondly, this article establishes an object association model based on greedy matching to overcome the correlation errors between objects caused by inaccurate detection and prediction. Finally, this article intends to study a multi-target tracking method with strong anti occlusion ability to achieve effective classification of different types of objects. The average tracking accuracy for dynamic occlusion (other vehicles, pedestrians) is 82%, and the average tracking loss rate is 13%. The research results of this article will provide new ideas and methods for solving object detection and tracking problems in complex environments.