The traditional traffic light systems are poor in flexibility and intelligencefor their fixed timing modes. In view of the above problems, a machine-vision-based traffic congestion evaluation system is presented to evaluate the current situation of traffic jams in this paper based on the collected video intelligent analysis and processing. Vehicle counting is firstly realized by HOG-feature analysis, AdaBoost method and RFID technology. Traffic states are evaluated in the Spark platform. The result of the experiments shows that our system can realize adjusted transformation time of traffic lights according to actual situation of the current traffic environment, then achieve the purpose of relieving traffic pressure dynamically.