NovelSense UG is concerned with the development of innovative sensor applications for the mobility of the future. In particular, new technologies for the infrastructure-side acquisition of traffic data are to be developed and tested. The goal is to develop hardware and software products for traffic planners and infrastructure developers. In the present project, AI methods will be used to enable continuous detection and counting of specific road users with a camera. A high application potential results from the current trends in the field of sustainable and multimodal mobility. Here, there is a considerable need for reliable traffic data on topics such as bicycle traffic and new mobility offers such as e-scooters.
A battery-operated and camera-based traffic counting device is being developed, which is characterized by its small size, low energy consumption and flexible installation options. A camera is used here, which can be mounted in a specially developed housing on the side of the road. The recognition of road users is based on the camera data in the device with a hardware-accelerated AI coprocessor using a neural network. The AI coprocessor allows classification and detection on real-time video data. Tracking and speed calculation of detected road users is then performed. The focus of the project within the Smart Data Innovation Challenge is the realization of the detection of e-scooters, which are already used en masse in cities like Berlin, Hamburg or Munich. Here, both the passage counting and the direction detection are to be realized.
The detection of e-scooters is achieved using transfer learning with feature extraction portions of pre-trained networks as well as fine-tuning the weights of newly added classification layers. The challenge is to optimize the deployed network to minimize the loss of detection accuracy while respecting the resource constraints of the hardware AI coprocessor, for example, current methods to reduce network quantization will also be investigated.
In the market for temporarily installable traffic counters, virtually no camera-based devices are currently deployed. The technologies used are mainly induction or pneumatic based technologies which have some disadvantages in terms of accuracy and range of application. This is mainly related to the differentiation of different road users. A particular advantage of the proposed technical solution is the privacy-compliant collection of data. In addition, the recognition of e-scooters is not currently offered by other providers on the market, regardless of the technology used. The data required for training, testing and validation of the neural network are compiled from freely available data sets as well as from data sets collected by NovelSense and can be used free of charge by follow-up projects.
The results from the project will be used directly in product development and will enable detection and counting of e-scooters in urban environments. To this end, test partners are currently already being sought and cooperation agreements concluded.
01.11.2019 – 30.04.2020
Yexu Zhou (firstname.lastname@example.org), KIT
Sascha Rudolph (email@example.com) NovelSense