Uber’s Autonomous Car Accident Investigation

Harveen Singh Chadha
Towards Data Science
5 min readMar 23, 2018

--

On March 19, one story shocked the autonomous vehicle industry when one of the uber’s self driving car ran over a woman in Tempe, Arizona. This was the first incident of loss of life after been hit by a self driving car. This seriously creates a serious doubt on all the claims that are put on the safety features provided by Self Driving Cars. On March 23, the Tempe policed released a video of the accident which puts many things into question. The full video can be viewed on youtube here.

Image Source (https://www.aol.com/article/news/2018/03/19/woman-dies-in-arizona-after-being-hit-by-uber-self-driving-car/23389700/)

Analyzing Dashcam Video

The full video can be viewed here.(Source- ABC Action News). Lets see the first image captured from the dashboard to analyze what all we get information.

5 seconds before collision.

There are a number of points to note from here-:
1. The lane lines are clearly visible.
2. There is car far away on the road and some lighting from the buildings.
3. The road has street lights!
4. With a naked eye, except a far car nothing is visible.

Apart from these details-
1. Car was driving at 38mph and the speed limit was 45mph. (Source )
2. Car was driving on a low beam.

2 seconds before the collision

Jumping to 0.07, 2 seconds before the collision, it can be observed that both on the left and right, we have street lights so there is sufficient light, but also on the left side there is a black portion of the road as well.

Left Street Light Passes by in the next frame of the same second

We are still in the seventh second and still no sign of any pedestrian, watch as the left street light passes by in this frame.

Still in the 7th second and both lights have passed. Now, It is the first time we see the footsteps of the pedestrian.

1 Second before the collision, it is sort of clear from the dashboard cam that the pedestrian is visible or there is some obstruction.

Well, we are in the 8th second and now we are able too see the lower half of the pedestrian and the bicycle.

Now, the woman is clearly visible with her bicycle. But did the car slowed down or even tried to slow down? No! Again an important point to note , the woman is wearing a black shirt/jacket. But what about the bicycle or the white shoes of the woman?

I have intentionally skipped the image for the 9th frame. Even at the 9th Second, the car did not slow down! That’s strange!!

Camera is not at question

It is true that the woman was observed about 50 to 60 frames before the impact, still the machine could not catch and slow down! But the point here is not about the camera, it was quite dark and the camera could miss that, also the upper half of the portion was black and could miss that too. The bicycle even was not visible fully. Now the main question comes. What happened to the LIDAR? LIDAR is one of the most expensive equipments used by self driving cars and it did not respond!! Anyone can argue the fact the Radar can miss some information but that is not the case with LIDAR. Did LIDAR treat the woman as noise? Well, it looks like this only, if this is it, we definitely got a use case to work on. One more point to ponder here is the woman didn’t even for a second was aware of the car!

The LIDAR didn’t work, ok we don’t have expectations from RADAR but that also didn’t work. The camera even after spotting the bicycle did not work. This is not acceptable and requires a detailed investigation, it needs to be understood what the LIDAR took this woman as and why was the input from the camera discarded. The fusion of all the sensors and Tech failed for 3 seconds and this is very surprising!! Also,The video footage is not enough!

The question of who is at fault still remains. Was the driver at fault ? It is clearly visible from the interior view that the driver was looking down just before the collision. It was her responsibility to control the car in any such circumstances but what I believe is that even a normal driver could not have avoided the accident and that is the exact reason why we are making self driving cars for, to avoid driver mistakes. The scene was pretty dark and unfortunately the woman was at the wrong place at the wrong time. If she would have crossed even from a place where the street lights had enough light, she might have been saved. But now the question goes, we have the footage we have analyzed it but who is at fault? Uber, Woman, Tech?

Well, one thing is pretty clear now, the self driving car industry needs rigorous testing to gain the trust of people back. Also, detailed investigations like we had in the case of an Air Crash Investigations on such incidents is required and should be shared amongst various players to raise the safety bars of this industry.

Please feel free to share your views.

--

--

Data Scientist| Currently building SOTA Automatic Speech Recognition Engine for Indic Languages | On a mission to change the world by motivating people |