Human:
My testing indicates that a 10ms delay was always too short and a 20ms delay was too short for close objects (which is counter-intuitive to me) - at least for my SR04 sensor.Increasing the delays in my larger code base reduced but did not eliminate my false negatives. Testing with my actual head while as stationary as I could make it produced lots of false negatives, depending on what it bounced off. In some cases, there were no false negatives. Moving my head while not walking was problematic, too. I suspect I will get more false readings while walking because this adds sensor vibration.
To sum up: Minimum delay between pings for my hardware was 40ms. I have some inherent false negatives due to the target shape, target motion, and sensor vibration in my application. I'm going to keep using my digital filter, unless I've missed some other solution. I'd be happy to share my filter with others if it's of general use.
Okay, so really the only problem is the delay between pings. I'm not suprised that a 10ms delay causes a lot of problems. I'm actually kind of suprised that I don't have a problem with a 10ms delay as that's really too quick (sensor is getting an echo from a previous ping). It could be the environment I'm in absorbs more sound or maybe my ping emitters are not as strong as yours. It also makes sense that a 20ms delay causes some problems, but not as many.
Because most of these ultrasonic sensors are designed to work up to 500cm away, that distance takes sound around 29ms to travel. So, 29ms between pings is the fastest you should really ping a sensor. If you ping faster, you risk getting an echo from a previous ping, causing the current ping to give a zero reading.
If we follow the baseline guide I've set for NewPing, and not delay less than 29ms between pings, my guess is that you will get basically no false negatives. The reason that 20ms gave better results with the trollmaker sketch with close objects with a 20ms delay is that the NewPing library has a timeout value for the initiation of the echo sensor while trollmaker using pulseIn can wait up to a full second. That timeout value in NewPing is the MAX_SENSOR_DELAY which is set to 18000us (18ms). This is plenty long, and it should never really be an issue as long as you keep the pings 29ms apart.
Set the delay to 29ms in the sketch and see if you get any false negatives. I would suspect you get basically none.
So the question is, were you all along trying to do pings too quickly which is why you concluded that the SR04 sensor gave many false negative readings? All of my sample sketches show delays in the 33 to 35ms range with comments not to go quicker than 29ms.
Tim