"Robotic household appliances" that free people from doing housework, "Work assist systems" for natural customer interaction, "Automated driving systems" to realize a safe mobile environment: Panasonic is conducting research to develop artificial intelligence (AI) technology that will assist people in a range of fields. An example of this is "Pedestrian Detection Technology based on Deep Learning," which aims to assist the human eye by utilizing Panasonic' s accumulated expertise in imaging technology.
Replacing the Human Eye
One of the key functions required to realize automated driving is the ability to detect pedestrians and obstacles that may pose a danger while the car is moving. Panasonic is utilizing Deep Learning, which automatically learns the features and patterns of a huge volume of data (over several hundred thousand files) and then recognizes and categorizes them, to newly develop High-Precision Pedestrian Detection Technology. This will enable the detection of pedestrians, even if they are partly obscured. Moreover, this developed technology uses an original algorithm that reduces the number of calculation and processing steps, which will enable real-time pedestrian detection.
Problems with Existing Pedestrian Detection Techniques
Up to now, mainstream pedestrian detection methods have relied on camera images. This involves using existing image recognition technology to identify regions on the screen with images that appear to be pedestrians (the pedestrian detection candidate region). An algorithm then repeatedly carries out successive detection processing on these identified regions using Deep Learning.
Flowchart for Existing Pedestrian Detection Algorithm
However, this type of algorithm has the following drawbacks.
- If there are numerous pedestrians detected on the screen, there will be an equivalent large number of potential pedestrian detection regions. This requires the same amount of repeated identification processing later on during the pedestrian detection processing stage. These necessary calculations take time to complete.
- If there are any regions left out in the initial stage of image recognition to identify potential pedestrian detection regions, the pedestrian detection process may not function properly later on.
- Extending the range of objects for potential detection beyond just pedestrians requires redesigning, for each object, the amount of pedestrian-like features to permit identification using image recognition.
Detecting 'Pedestrian-Likeness' by Single Identification Processing
Panasonic' s newly developed technology features our own algorithm that can calculate pedestrian-likeness by carrying out identification processing only once for each image. The key to this is a new algorithm that uses Deep Learning to plot a likelihood map (a probability distribution map that distinguishes pedestrians from other general scenery) for displaying pedestrian-likeness in each image region.
Using this algorithm, and based on the plotted likelihood map, it is possible to reliably detect pedestrians in various postures and situations, such as when using an umbrella.
Flowchart Showing our Developed Pedestrian Detection Algorithm
Possibilities for Expanding the Range of Detectable Objects
The newly developed algorithm also makes it possible to easily expand the range of detectable objects to beyond pedestrians. Using image regions based on Deep Learning, only a single image is required to carry out identification processing for creating a likelihood map. This eliminates the need for a massive increase in the number of calculations, and makes it possible to add other non-pedestrian objects for identification processing, such as cars, roads and signs.