Internet congestion existing solutions such as active queue management algorithms have many shortcomings, mainly related to the detection phase. These algorithms depend on routers' buffer statistics to detect congestion and their performance is highly affected by the environment and the parameters that are used. In this paper we are proposing a mechanism that is capable of detecting congestions by monitoring passively an aggregation link. The proposed mechanism does not need parameterizations since all the used parameters are deduced from public real internet traces using statistical approaches and uses TCP delays as a detection parameter. It is dynamic since the detection is proportional to the severity of the congestion. Experimental results have shown that the proposed mechanism is able to detect congestion rapidly and does not suffer from false alarms.