Abstract: To solve the problem of missing vehicle detection resulted from vehicle occlusion in heavy traffic congestion using video processing technology, a method of robust vehicle detection and estimation of the lane-by-lane arrival cumulative curves for traffic congestion was presented. This method included three aspects:firstly, non-congested areas were detected to avoid the meaningless work of detecting the stopped vehicles with occlusion within traffic congestion. Then, the robust results of vehicle detection were obtained by fusing the AdaBoost vehicle classifier and under vehicle shadow detection based on the hypothesis generation and verification framework. Finally, the lane-by-lane arrival cumulative curves was estimated accurately, after each vehicle was classified into the specific lane by using the stable features without projection distortion, which resulted in the effective analysis with traffic parameters. The experimental results show that this real-time method can detect vehicles robustly and obtain the traffic parameters accurately during the peak period of traffic congestion. It is thus clear that this method can effectively avoid complex implementation dealing with vehicle occlusion, and is significant to solving the practical problems of high cost, heavy workload and many uncertain factors when investigating vehicle arrival rate and headway data.