Detection Algorithm Improvements for Enhanced Traffic Flow and Safety – Part 1: Vehicle Detection

In the first of his three-part blog series that will explain how improvements to detection algorithms are enhancing traffic flow and safety at the intersection, Iteris’ Michael Whiting focuses on vehicle detection, and the major detection issues and solutions that brought about these life-saving improvements.

When I joined Iteris over 13 years ago, I was somewhat oblivious to all the magic that goes into running a traffic network in a city. The signals just know that you are there, right? My naivety extended to the fact that people, including myself, would often complain about not getting a green light, whether or not we had waited a reasonable amount of time.

While some traffic signals are on a fixed time, others rely on detectors to accurately and consistently detect a vehicle’s presence to actuate the signal. Unfortunately, as I have learned in my time in the industry, vehicle detection issues can significantly affect traffic flow and human behavior.

A breakdown of common detection issues

Detection issues fall into two broad categories – missed detections and false calls – both of which can pose a considerable safety risk. For example, running a red light out of frustration after waiting a reasonable amount of time and not being detected can lead to three possible outcomes, each more severe than the last:

1. Nothing happens; you go on your way
   
2. You receive a ticket for running a red light
   
3. You collide with a vehicle, pedestrian or cyclist
   

The second category of detection failure is the false call, which involves sending a signal to the traffic controller to change the light when no vehicle is present. From a traffic engineering perspective, this creates an efficiency problem as it impedes traffic flow. From a commuter perspective, it can lead to frustration: “Why am I waiting at a red light when there is no other traffic around!”

Once again, frustration can lead to poor or aggressive driving decisions and heightened safety risk. Research conducted by the AAA Foundation found that 55.7% of fatal crashes involved at least one driver who was reported to have performed at least one potentially aggressive action.

Improved vehicle detection algorithms for safer streets

So what does this all mean to Iteris detection systems and the algorithms at the heart of their performance? Whether we conscientiously knew it or not, we increased the safety at every intersection our sensors were deployed at as we continuously developed and enhanced our detection systems’ accuracy and performance.

Three major vehicle detection algorithm improvements were responsible for this markedly improved performance:

1. Eliminating shadows for accurate detection

We started with the notion of "if something is there, then it is probably a vehicle to, solving significant detection challenges with shadows. Shadows cast by moving vehicles in adjacent lanes or complex, slow-moving shadows from buildings, trees and other road furniture impair vehicle detection, and often cause false calls.

One of the biggest technical challenges is making a machine think like a human. When we look at an image, we naturally identify and classify what we see. Current artificial intelligence efforts attempt to recreate this human trait, which I’ll talk about more in a future blog post.

There are more traditional machine vision techniques at work here; frame differencing and edge detection are two examples. To overcome shadow issues, we use both of these techniques. By attaching a shadow cast to a recognized vehicle, we can ignore or reject this as part of our detection process. Most shadows cast by natural objects rarely have a sharp edge, which helps us differentiate a shadow from a vehicle.

2. Accurate detection in various environments

The most critical improvement to our algorithm has been to overcome detection issues in the most adverse conditions, including rain, snowstorms, nighttime and low lighting caused by fog (see Figure 1).

Figure 1: Vehicle detection through heavy fog.

When we look at a low-contrast scene from a machine vision viewpoint, what does the machine see? It sees a very uniform grayscale with some small peaks where vehicles may be (see Figure 2). The algorithm focuses on these peaks and looks for key features that would be related to a vehicle – front grills, windscreens and headlights, for example. Linking these features together provides high statistical confidence that a vehicle is present.

Figure 2: Grayscale view in algorithm.

3. Self-assessing and correcting algorithm

Overriding all this is the detection system’s ongoing self-assessment of performance and image quality. We have built many fail-safes into our detection products to provide data outputs to the traffic controller anytime performance is compromised. The engineering team at Iteris is proud of the advanced, high-quality systems we produce, knowing that any intersection where they are deployed will enable the most efficient traffic flow and increase the traveling public’s safety.

In part two of this blog series, which aims to explain how improvements to Iteris detection algorithms are enhancing traffic flow and safety, I will focus on how our algorithms detect and differentiate pedestrians and bicycles.