Queue Length Estimation at Junctions

Mohammad Shokrolah shirazi and Brendan Morris

Proceeding of the IEEE International Conference on Advanced Video and Signal Based Surveillance, pp. 1-6, August, 2015, Karlsruhe, Germany.

Abstract

This paper presents a tracking method for vision-based queue analysis at junctions including queue length and waiting time estimation of vehicles. The tracking method works based on improving the optical flow method to track small and low quality vehicles with overlong waiting time. The improvement process is performed to keep track of overlong stopped vehicles with high level of robustness against occlusion by crossing pedestrians. The results of experiments are presented by estimating queue length, waiting time distribution and number of waiting vehicles for the highly cluttered video of a Las Vegas junction. The accuracy of the system is evaluated by comparing the queue analysis results with the ground truth.

Goal

Developing a system to estimate queue length using cameras which is an important parameter in control models to improve the passing capacity. Moreover, queue length estimation and associated delay (i.e. waiting time) are useful for devising traffic management strategies that would help to optimize traffic signals and improve the performance of a traffic network. There are non-tracking based methods that determine queue length based on introduced features on road by vehicles such as Local Binary Pattern (LBP) , spatial edges , FFT , and image gradients. However, the paper presents the tracking-based method which provide more traffic parameters such as number of waiting vehicles and waiting times which can not be obtained by non-tracking based methods.

Vehicle Detection & Tracking System

The system detects vehicles using back ground subtraction at first and using bipartite graph to initialize tracks. Vehicle tracking system benefits cooperation of bipartite

graph with optical flow to track detected vehicles. The vehicle tracking system is shown below. Detected vehicles in the motion area are given to the tracking system which uses bipartite graph at first to track detected vehicles for 3 frames and initialize the tracks. The initialized tracks use optical flow during the tracking phase.

Queue Analysis

Vehicles paths are recognized using temporal alignment techniques for similarity measures of trajectories with typical paths which have been collected for each lane. Longest common subsequence (LCSS) distance is a popular technique for comparing unequal length trajectories. Tracks are labelled regarding each lane and their moving or waiting state is determined at each frame. When the waiting state of vehicles is determined, they become candidates for queue length estimation if there is spatial proximity between them. The track candidates are saved into separate lists regarding each lane for queue length estimation.

Queue length is gauged for waiting vehicles of each lane using feature points of the tracks. Texture and corner feature points of stopped vehicles are used to estimate the queue length. Since vehicles’ queue lines might have different orientations, the line is projected into each lane by selecting (x, y) coordinates of feature points according to highest and least y values. Two selected feature points find the nearest neighbor coordinate from typical paths and the Euclidean distance is used to measure the distance. Figure above shows an example of queue length estimation. The queue length estimation with red line have been estimated from the feature points (green points) of stopped vehicles shown by aqua bounding box.

Results

The proposed system was implemented in C++ using OpenCV and it was evaluated for a highly cluttered video of the Las Vegas junction. The experiments included evaluation of the system for queue length estimation, count of waiting vehicles in queue and waiting time estimation of vehicles in regards to different lanes.

Figures above show the queue length estimation and corresponding number of waiting vehicles regarding two different lanes. They show a similar pattern and traffic signal phases can be inferred from them. Different queue lengths for the same number of waiting vehicles are shown (e.g. 24-50 seconds) due to tracking error and different sizes of waiting vehicles in lane 1.

Video