Thursday, June 28, 2012

Detecting People on Bikes at Traffic Signals - Radar or Video Detection

Initial findings from Comparison

In the City of Portland, actuated traffic control is primarily detected by inductive loop detectors. We have used video detection, radar, and other emerging devices sparingly, because we have inductive loops as a tried and true method for detecting all users. There has also been significant research that raised concerns about detection. The City has also worked hard to maintain inductive loops and there is significant infrastructure that has resulted in effective results.

In this study (with limited number of samples) conducted by a Portland State University intern we compared video with radar (presence detection type) to determine the effectiveness of each device. Video detection detects vehicles passing through a “field-of-view” and then is communicated to a processor that allows the sensor data to be accessed by the software.

The radar device tested is a device that generates 16 separate radar beams in close proximity to create an effective field of view. According to the vendor, the sensor detects each vehicle within its field of view, remembers their positions and can predict future movements for "greatly improved detection accuracy".

BACKGROUND

The Broadway Bridge is located on the Willamette River in Portland. The intersection of Broadway and NW Lovejoy is located on the west end of the bridge. Of the westbound motor vehicle lanes, the right lane is a right turn only onto NW Lovejoy, which leads to the Pearl Distinct. The left lane is a left turn only onto NW Broadway which leads to Downtown Portland.

The bike lane traveling west is on the north sidewalk of the Broadway Bridge. The bicycle lane then splits into left and right turn lanes on the west end of the bridge separate from pedestrians. The left turn bike lane has its own signal and signal timing. The left turn bike lane and the two motor vehicle lane signals are activated with the aid of the detection for each approach.

DATA COLLECTION

Both forms of detection are installed at the intersection of Broadway and NW Lovejoy on the west side of the Broadway Bridge. The video captures the westbound lanes coming from the Broadway Bridge.

The effectiveness of the detection of cyclists with the two types of systems, Radar and Video, were evaluated for the left turn bike lane. When each system responds to a moving vehicle within its detection zone the names of the detection devices, “Radar” and “Video”, are displayed on the video. When the detection devices are functioning properly, each vehicle traveling within the detection zone is detected and added to the signal timing.

Video was recorded on March 14, 2011 at the intersection. These two videos were watched to determine the effectiveness of each system at detecting cyclists. If the observer of the video could see the person on the bicycle and there was no detection it was recorded as missing the detection. We studied this during both night and daylight conditions. The videos were recorded from 5-7 AM, before daylight, and from 8-10AM during daylight hours. The weather conditions during both recordings were wet and rainy.

The time and number of cyclists entering the left bike lane was recorded, the time of the activation of the video and the radar detection were also recorded. A successful detection is when either the detector came on and remained on while the cyclist entered the detection zone or when the detector pulsed on and off when the cyclist passed the sensor and the system then added the bike signal to the timing cycle. A failure was when a detector failed to recognize a cyclist (a missed call) on the left bike lane. The accuracy of the collected data was in seconds.

FINDINGS

From watching the video recordings there were a total of 205 cyclists passing through the left turn bicycle lane at the intersection during the 4 hours observed. 59 cyclists were counted during the 5:00 AM to 7:00 AM video, and 146 cyclists were counted during the 8:00 AM to 10:00 AM video. These cyclists were visually counted as a basis to test the Radar and Video units’ accuracy.



The Video detected 100 percent of bikes during the pre-daylight hours and was 99.3 percent during daylight hours. The Radar only detected 91 percent of cyclists during pre- daylight hours and 96 percent during daylight hours. See Table 1. The Video was susceptible to false calls and Radar often missed calls. During our limited observations, there seemed to be no difference in detection during evening hours between bicycles with flashing lights, steady lights or no bike lights at all.

PART 2: RECALIBRATION OF RADAR

The radar vehicle detection unit was recalibrated and retested in an attempt to increase the accuracy of bike detection.

A new video of the intersection was recorded on May 5, 2011 from 6:50 AM to 9:00 AM. Data was collected from the video in order to determine the accuracy of each of the detection devices. The time and length of activation of the Video and the Radar, and the time and number of cyclists that entered the left bike lane was recorded. A successful detection is when either the detector came on and remained on while the cyclist entered the detection zone or when the detector pulsed on and off when the cyclist passed the sensor and the system then added the bike signal to the timing cycle. A failure was when a detector failed to recognize a cyclist (a missed call) on the left bike lane. The accuracy of the collected data was in seconds.

Unlike the previous video data collection in Part I, data was collected to also include the number of cyclists that ran a red light.

During the time of this video recording, construction work was taking place on the Broadway Bridge. As part of the construction on the bridge, the sidewalk on the south side of the bridge was closed. Cyclists coming from the south on Broadway Boulevard were crossing over the vehicle lanes to access the sidewalk on the north side of the bridge. These cyclists were using the left bike lane to access the sidewalk on the north side of the bridge. These cyclists were not detected by either detection system.

Detections were placed into three different categories:

Bikes that approached the intersection on a red and stopped at the intersection
Detection of bikes that approached the intersection during a green light and traveled through
Bikes that ran a red light

Each of these detection categories had four sub categories:

Missed detection by Radar only
Missed detection by Video only
Detection of a bike missed by both detectors
Detection by both detectors


FINDINGS

A total of 245 cyclists were recorded within the two hour period.

Of all bikes that approached the intersection from the left bike lane, the Video detected 100 percent of bikes. Radar detected 98.8 percent.




22 people on bikes (nine percent) did not comply with the red bike signal.

Of cyclists that were able to cycle through the intersection without slowing down on a green bike signal, Radar detected 82 percent of these people on bicycles, whereas Video detected 99.6 percent.

2 comments:

Glenn Grayson said...

Peter, I'm a bit surprised that this interesting finding didn't generate any comments from your readers. I think the analysis and results are very useful, and probably already have helped to formulate or solidify your opinions about detecting bicycles with non-pavement-invasive detection systems. A few follow-up questions: (1) the blog subtitle notes that these are initial findings - has a final report been prepared? (2) Was this study done using Autoscope and Wavetronix products? (3) Because technology has advanced significantly since your study was undertaken, might you be thinking about repeating the experiment with the new wireless bicycle-specific detection products from FLIR and Iteris?

pkoonce said...

Glenn-
We're working with interns and as a City employee, we aren't always producing reports with final findings due to the interns coming and going.
I deliberately stayed away from product names because the rigors of the study were not like what a proper research study (see work from Purdue University) would do.
We do have studies underway with Flir and Sensys.