This project aimed to detect, track, differentiate, and count different brands of the scooter (e.g. Lime, Flamingo) in public areas in real-time, on inexpensive hardware.
Detection, tracking, and counting had to be achieved at the local device level. A more conventional approach would be to send field video data to a central server for batch processing and analytics. This has several downsides:
➜ Typically much higher overall costs.
➜ Video requires high bandwidth.
➜ Privacy concerns from storing personally identifiable images centrally.
➜ No access to real-time data or analytics that is important for rapid decision making.
The availability of cheap, low power compute boards like Raspberry Pi, and new hardware AI accelerator modules (in this case Google’s Coral inference engine) meant we could complete all computer vision and AI tasks on the field device.
This local processing means we only need to send small, anonymised data packets to a server – in this case the numbers of different scooter brands crossing an area of pavement in a given time period.
The same process can easily be applied to any type of traffic – pedestrians, cyclists, cars, buses (or any mixtures of these), with data made available in real-time to upstream systems like traffic control centres.
Intranel was able to apply our model training processes to detect multiple types of scooters with high accuracy, with only a small sample set. Training data came from hiring a set of scooters for a few hours and taking videos from multiple angles, from which still images were extracted.
We were able to apply our hardware-accelerated tracking algorithms to ensure even higher recognition rates as the same object is tested multiple times, with a majority vote applied on the detected scooter type.