Transportation perception

SmartParking

Which object detector best supports the mixed visual demands of a smart parking environment?

Presented at IEEE ICTC 2022 in Jeju Island, South Korea

SmartParking repository TraPedesVeh dataset IEEE paper

Overview

How do Faster R-CNN and SSD detector families trade detection quality against training speed when they are evaluated on the same multi-class parking dataset?

SmartParking is a computer-vision study of joint vehicle, pedestrian, cyclist, bicycle, bus, and traffic-sign detection. Unlike work that treats each object family as a separate problem, this project evaluated the combined perception task required around a parking facility.

The study also introduced TraPedesVeh, a labeled mini-dataset assembled to address the limited availability of images containing the multiple road users and signs that a parking system must understand at the same time.

Smart parking object-detection system from dataset preparation to traffic-scene deployment

The complete system connects the TraPedesVeh dataset to preprocessing, feature extraction, classification, evaluation, and real-traffic testing for vehicles, pedestrians, and traffic signs.

Methods

TraPedesVeh dataset

Images containing vehicles, pedestrians, cyclists, bicycles, buses, and traffic signs were collected, annotated, normalized, and converted into a common detection format.

Six detector comparison

The study evaluated SSD and Faster R-CNN variants with MobileNet, ResNet, and Inception-ResNet backbones under a shared computational setting.

Multi-metric evaluation

Detection accuracy, average precision, average recall, training time, classification loss, localization loss, and total loss exposed the speed-accuracy tradeoff.

91.5%best average detection accuracy

6detector configurations evaluated

6reported object classes

5.636 sfastest reported training time

Contributions

A unified evaluation of vehicle, pedestrian, cyclist, bicycle, bus, and traffic-sign detection.
TraPedesVeh, a labeled mini-dataset for intelligent transportation research.
Faster R-CNN with Inception-ResNet achieved the highest average detection accuracy at 91.5 percent.
SSD-MobileNet at 320 x 320 trained fastest, documenting the practical speed-accuracy tradeoff.
A reproducible baseline for perception around smart parking environments.

TraPedesVeh dataset construction and annotation process — The dataset pipeline moves from class-keyword selection and image retrieval through screening, annotation, TFRecord conversion, and label-map creation.

Interpretation

The highest-accuracy model was also the slowest to train, while the fastest SSD configuration sacrificed substantial detection quality. That tension is central to smart-parking deployment: the best model depends on whether the system prioritizes detection fidelity, limited hardware, retraining speed, or real-time response.

Current scope

SmartParking is a perception and detector-benchmarking project. It is separate from PANDA, which forecasts future parking occupancy and synchronizes predictions with a Cesium digital twin.

Additional figures

SmartParking detection result for pedestrians and a vehicle — A qualitative result showing joint pedestrian and vehicle detections in a street scene. These mixed scenes reflect the perception demands near parking facilities.

SmartParking detection result for pedestrians, cyclists, and bicycles — A multi-object result with pedestrians, cyclists, and bicycles detected in the same urban scene.

SmartParking comparison result for vehicles and pedestrians — An additional detector output used to compare localization and class confidence across the evaluated model families.

SmartParking result detecting pedestrians, vehicles, and a traffic sign — This dense urban scene tests simultaneous localization of pedestrians, vehicles, and a traffic sign, including partially overlapping people and small distant objects.

SmartParking result detecting pedestrians and bicycles in a crowded street — A crowded street result demonstrates multi-class perception with pedestrians and bicycles at different scales, orientations, and levels of occlusion.