Analyzing visitor behaviors in the Linnaean Botanical Garden,UppsalaTowards a privacy-preserving camera node

Abstract: Natural environments are good for people’s health. The LinnaeanBotanical Garden in Uppsala provides different offerings like alpinerocks, flower beds and tropical houses to visitors. We are interested in the visitors’ activities and their interactions with the gardenfrom data-based perspectives. As collecting activity data will involve privacy concerns, a privacy-preserving sensing platform needs to be developed and we focused on the face-locating and masking approach. A testbed centered on Raspberry PI (RPI) 4 is built and a gardenvisitingvideo is recorded with the help of volunteering visitors. On the RPI we implemented a CPU-based face-locating pipeline by combiningmotion detection (MOG2), frame-skipping, HOG+SVM head-shoulderdetection, simple CNN-based double checking and MOSSE tracking. The pipeline achieved 37% AP and 17 FPS processing speed (using frameskipping) on the 15 FPS field test video. Additionally, a proof-ofconceptactivity visualizer was also made, which projects visitors’positions to bird view and adds up their histories as heatmaps thatcan be overlaid on aero maps. The current face-locating pipeline still shows considerable space for accuracy improvement. Further approaches like Deep-Learning object detectors on the testbed may also be feasible with the aid of accelerated edge computing devices. As the heatmap basically reflectsthe visitor’s flow, using professional analysis software, more behaviors may be discovered. This thesis provides some first insights on how to develop a privacy-preserving camera node, which would help the design of future gardens.