1.1 Motivation
For typical driving experience in the urban areas in Taiwan, there is a high possibility that a driver need to stop at the traffic signals for more than 60 seconds. If a driver is distracted from the waiting for the traffic signal to turn green, e.g., for tuning for a radio program or finding accessories in the purse, he or she is very likely to miss the traffic signal transition, or be honked by the rear vehicle for blocking the traffic. We want to develop a detecting system which starts to monitor traffic signals and vehicles when the driver stops at a red traffic light, and notifies the driver when it is time to move again.
As car video recorders become popular, more and more people attach these devices to the windshield or on the dashboard to record their journeys. Compared to the radar or laser solutions, camera-based solutions are more accessible. Many brands of car video recorders, ranging from NTD 3,000 to NTD 10,000, can be easily bought. Furthermore, considering the computation power of mobile devices today, e.g., HTC One X with 1.5 GHz quad-core CPU (2012), car video recorders with more powerful computing capabilities are also expectable. Therefore, we choose vision-based approach to detect traffic events and issue timely notifications to the driver.
1.2 Review of Related Works
While driving on the roads, the traffic conditions may change rapidly. Besides controlling the vehicle, a driver has to continuously attend to and react immediately and properly to receive visual and acoustic information around the vehicle. However, humans may not maintain a high level of concentration for a prolonged period of time, nor has automatic vehicle driving not been mature, the development about Driver Assistance Systems (DAS) [1] becomes an
2
important research field. A DAS monitors surrounding environment continuously by sensors and provides extra information to the driver. Besides assisting the driver, some DAS technologies can be the foundations of autonomous driving in the future.
Categorized by applications, many technologies [2] have been developed for DAS which include: lane departure warning system [3] [4], adaptive cruise control [5], collision avoidance system [6], pedestrian protection system [7] blind spot detection [8], driver drowsiness detection [9], automatic parking [10], traffic sign recognition [11] [12] [13], etc. Some of these technologies have also been commercialized. For example, Volvo’s City Safety [14] can prevent collision at speed under 50 km/h by using an active sensor to scan the region 10 meters in front of the vehicle, and will apply brake if the driver is not aware of the obstacle.
While most DAS technologies attempt to increase car safety by informing the driver immediately when the system detects a possible danger, others, such as automatic parking and traffic sign recognition intent to provide useful and convenient functionalities to ease the driving effort. This thesis will concentrate on the convenience point of view.
As described previously we plan to design a system to notify a driver stopped by the traffic signal or stopping vehicles in front when to start moving again. In general, whether a notification should be issued will be determined by the transition of traffic signals as well as the behavior of front vehicles. Color and shape are important features for detecting traffic signals. Park and Jeong [15] find traffic signal regions by thresholding RGB values, applying circularity check and rejecting candidates which do not last long enough. Gonzâles et al. [13]
use Hough transform and aspect ratio to extract possible traffic sign locations. Meanwhile, multi-frame validation is proved to be effective to reject unstable candidates.
As for vehicle detection, Zheng and Liang [16] train classifiers by using RealBoost [17]
with edge-like stripe features and deal multi-view problem with Cluster Boosting Tree [18]
algorithm. Kuo and Nevatia [19] detect multi-view vehicles with a tree structure detector in
3
which each node is a Gentle AdaBoost classifier [17] based on HOG feature [20]. Chen and Lin [21] identify vehicles at night by analyzing candidates of brake light regions in the frequency domain. Jazayeri et al. [22] represent the movement of vehicles in a spatiotemporal image constructed by horizontal edges. To the best of our knowledge, there are no studies about detecting the transitions of traffic signals, which motivates the proposed approach.
1.3 Overview of Proposed Methods
In this thesis, we design an innovative and practical daytime DAS system, entitled Time2Go, which is capable of reminding the driver when it is time to drive forward by analyzing
the video clips captured by a car video recorder. Fig. 1 illustrates the flowchart of the Time2Go System which is initiated whenever a video frame is captured. After it is converted to HSV space the frame is then sent to different modules in the system depending on the driver is moving on the road (state MOVING) or stopping at the traffic signal (state STOPPED), as determined by the Ego-Motion Detector. (Both a spatiotemporal-profile-based method [22] and a scan-line-based method will be considered for the Ego-Motion Detector in this thesis.)
If the system determines that driver’s vehicle is stopped, video frames will be sent to the Vehicle Detector and the Traffic Signal Detector in an interleaving way. The Vehicle Detector finds the locations and motions of front vehicles based on Gentle AdaBoost [17] and HOG features [20]. The performance is further improved with a red-region-pair checking. The Traffic Signal Detector is used to locate the candidates of traffic signals, and sends out a notification when a red-to-green happens. The concept of multi-frame validation is widely used in all three modules.
4
Vehicle Detector Traffic Signal Detector Frame
5
1.4 Thesis Organization
The remainder of this thesis is organized as follows. Chapter 2 elaborates the detail of each modules of the proposed DAS system, including Ego-Motion Detector, Vehicle Detector and Traffic Signal Detector. Dataset collection, experiment results and assessment of system performance are described in Chapter 3. Finally, conclusions and future works of this thesis are given in Chapter 4.
6