In 2020, about 9.8% of the German population rode their bike daily, and almost 25% did at least once per week. Due to lack of an adequate protection such as an external shield and the inability to respond in vehicular collisions, cyclists, as well as pedestrians and motorcyclists, are called Vulnerable Road Users (VRU). As not only the number of VRU, but also the number of their deaths continues to climb, reaching 350.000 in 2016, many regulations have been implemented and different measures to improve traffic safety have been developed (i.e. in-vehicle systems, using sensors as radar, infrared, camera or LiDAR or V2P technologies). However, these products cannot handle situation without line-of-sight (LOS). Solutions to increase safety at intersections and crossing have been developed, installing and integrating sensors in the road infrastructure. Information and data gathered by sensors, are then sent to a control unit, that estimates the trajectories of the different road users, to make a risk assessment of potential collision. In case of crucial situation, not only the vehicle but also the VRU will be warned. In this context, furthermore, trajectory prediction plays an important role, to foresee vehicles’ maneuvers and lane changing before approaching an intersection, to buy time in VRUs warning in case of possible a potential critical situations. The "People Mover" project, promoted by the city of Regensburg (Germany), aims to reduce the incidents where VRU are involved, installing a road infrastracure, in a usually crowded crossroad in the Business Park. This study, conducted within the trajectory prediction and collision warning module of the ”People Mover” Project, exploits offline generated physical map information, such as the pre-defined routes, to offer a real-time efficient system. A predefined route is a pathway, that can be traveled by a vehicle, following always the same lane, without any overtaking or changing lane maneuver, in accordance to traffic rules. Basing on the state of the traffic object (position, linear velocity and acceleration, heading angle and angular velocity), the trajectory is generated along the predefined route. The trajectory prediction module makes use of the maneuver recognition system: an IMM, Interactive Multiple Model filter, integrated with the information coming from the offline generated map information, establishes the current maneuver of a vehicle from the set of maneuver {KeepLane, ChangeLane, Curving}. In case of a ChangeLane, the path generated in order to pass from the current route to the target one is a piecewise quadratic Bèzier curve. On the other hand, pedestrians are not assigned to a route, and their path is computed by means of a simple CA (Constant Acceleration) motion model, with the current state as input. The collision prediction algorithm intersects then all the trajectories to detect a conflict area, area occupied by two objects at the same time, that would determine a warning for the VRU. The system, tested and validated using Matlab/Simulink and IPG CarMaker, has demostrated a good response time in the maneuver identification (delay less than 0.25s). While acceptable errors (between 3-5m have) been reported for trajectory predictions without abrupt movements in a time window of [0.5s, 3s], the error reaches 15-20m for fast acceleration/deceleration and/or for a longer time horizont. Although the system has detected all the collisions, demonstrating the crucial role of the lane change detection module, saving time in the risk assessment, many warnings have been raised because of the safety distance, because of the uncertainties integrated in the trajection prediction (the near in the future, the smaller the uncertainty). A good compromise would be to avoid prediction for long time horizons.
Un numero sempre maggiore di persone rinunciano all’utilizzo di un'auto, prediligendo spostamenti tramite bicicletta o a piedi. Per questo motivo, é stata introdotta nel Codice della Strada nel 2021 una nuova categoria denominata VRU, Utenti Vulnerabili della Strada, nei cui confronti sono previste misure per agevolare gli spostamenti e ridurre i disagi. Tali provvedimenti, seppur riducendo notevolmente il numero di incidenti in cui i VRU sono coinvolti, non risolvono il problema alla radice. Vari sistemi (sensori LiDAR e camere, tecnologie V2P) sono stati sviluppati negli ultimi anni per essere installati nei veicoli o nelle infrastrutture stradali, in modo tale da aumentare la sicurezza dei VRU e diminuire incidenti in situazioni di NLoS (Non-line-of-Sight). In tale cornice si colloca il progetto People Mover, promosso e finanziato dalla cittá di Ratisbona (Germania), volto a salvaguardare i VRU. Tale progetto, in fase di decollo, mira all’installazione di una infrastruttura stradale nella zona industriale, e precisamente in uno dei crocevia più trafficati. Lo studio condotto in questa tesi propone un sistema preposto al riconoscimento di rischi per i VRU, basato sulla conoscenza a priori della mappa stradale circostante e su un rilevamento precoce delle intenzioni di manovra dei singoli veicoli finalizzato ad una accurata predizione delle traiettorie. Il sistema generale é formato da tre moduli principali: il modulo di riconoscimento manovra, il modulo di predizione della traiettoria e quello di identificazione dei pericoli. La conoscenza pregressa circa l'ambiente circostante (predefined routes) permette il ricorso a metodologie computazionalmente non pesanti e a semplici funzioni, che garantiscono buone prestazioni a real time. Con il termine route predefinita si intende una traiettoria percorribile da veicoli (non pedoni), nel rispetto delle regole stradali, lungo le quali le future posizioni dei singoli soggetti non pedoni vengono predette. La definizione delle traiettorie avviene in base al modulo di riconoscimento manovra, il quale, attraverso un filtro IMM (Interacting Multiple Model) e le informazioni proventienti dalla mappa stradale, determina quale delle manovre dello specifico set {ChangeLane, KeepLane, Curving} é in atto. Nel caso di cambiamento di corsia, il passaggio da una route alla route obiettivo viene calcolato ricorrendo ad una doppia curva quadratica di Bèzier, che assicura un basso carico computazionale. Le traiettorie dei pedoni, indipendenti dalle singole route, si basano su un semplice modello di moto CA (Constant Acceleration). L'algoritmo binario di predizione delle collisioni interseca quindi le varie traiettorie per identificare la cosiddetta area di conflitto, area occupata da due soggetti nello stesso istante di tempo, la quale stabilisce l'effettivo rischio per il VRU. Tale studio, validato in simulazione tramite l'utilizzo di Matlab/Simulink e IPG CarMaker, ha dimostrato un buon tempo di risposta nell'identificazione delle manovra (ritardo dall'inizio della manovra minore di 0,25s). Mentre errori accettabili (3-5m) sono stati riscontrati nella predizione delle traiettorie per tutti i soggetti non sottoposti a cambiamenti di stato repentini in una finestra temporale di [0.5s, 3s] (independentemente dalle manovre), l'errore raggiunge i 15-20m per orizzonti temporali pari a 4s e/o in presenza di accelerazioni/decelerazioni improvvise. Sebbene il sistema abbia riconosciuto tutte le situazioni di pericolo con buon anticipo (2-3s), dimostrando la cruciale importanza del modulo di riconoscimento di cambio corsia, sono stati rilevati numerosi falsi positivi, causati dalle incertezze sulla posizione integrate nel module di predizione della traiettoria. Un buon compromesso risulta quindi limitare l’orizzonte temporale a t = 3s.
Design e simulazione di un sistema per predire collisioni tra utenti vulnerabili della strada (VRU) e veicoli.
LEONORI, CHIARA
2020/2021
Abstract
In 2020, about 9.8% of the German population rode their bike daily, and almost 25% did at least once per week. Due to lack of an adequate protection such as an external shield and the inability to respond in vehicular collisions, cyclists, as well as pedestrians and motorcyclists, are called Vulnerable Road Users (VRU). As not only the number of VRU, but also the number of their deaths continues to climb, reaching 350.000 in 2016, many regulations have been implemented and different measures to improve traffic safety have been developed (i.e. in-vehicle systems, using sensors as radar, infrared, camera or LiDAR or V2P technologies). However, these products cannot handle situation without line-of-sight (LOS). Solutions to increase safety at intersections and crossing have been developed, installing and integrating sensors in the road infrastructure. Information and data gathered by sensors, are then sent to a control unit, that estimates the trajectories of the different road users, to make a risk assessment of potential collision. In case of crucial situation, not only the vehicle but also the VRU will be warned. In this context, furthermore, trajectory prediction plays an important role, to foresee vehicles’ maneuvers and lane changing before approaching an intersection, to buy time in VRUs warning in case of possible a potential critical situations. The "People Mover" project, promoted by the city of Regensburg (Germany), aims to reduce the incidents where VRU are involved, installing a road infrastracure, in a usually crowded crossroad in the Business Park. This study, conducted within the trajectory prediction and collision warning module of the ”People Mover” Project, exploits offline generated physical map information, such as the pre-defined routes, to offer a real-time efficient system. A predefined route is a pathway, that can be traveled by a vehicle, following always the same lane, without any overtaking or changing lane maneuver, in accordance to traffic rules. Basing on the state of the traffic object (position, linear velocity and acceleration, heading angle and angular velocity), the trajectory is generated along the predefined route. The trajectory prediction module makes use of the maneuver recognition system: an IMM, Interactive Multiple Model filter, integrated with the information coming from the offline generated map information, establishes the current maneuver of a vehicle from the set of maneuver {KeepLane, ChangeLane, Curving}. In case of a ChangeLane, the path generated in order to pass from the current route to the target one is a piecewise quadratic Bèzier curve. On the other hand, pedestrians are not assigned to a route, and their path is computed by means of a simple CA (Constant Acceleration) motion model, with the current state as input. The collision prediction algorithm intersects then all the trajectories to detect a conflict area, area occupied by two objects at the same time, that would determine a warning for the VRU. The system, tested and validated using Matlab/Simulink and IPG CarMaker, has demostrated a good response time in the maneuver identification (delay less than 0.25s). While acceptable errors (between 3-5m have) been reported for trajectory predictions without abrupt movements in a time window of [0.5s, 3s], the error reaches 15-20m for fast acceleration/deceleration and/or for a longer time horizont. Although the system has detected all the collisions, demonstrating the crucial role of the lane change detection module, saving time in the risk assessment, many warnings have been raised because of the safety distance, because of the uncertainties integrated in the trajection prediction (the near in the future, the smaller the uncertainty). A good compromise would be to avoid prediction for long time horizons.| File | Dimensione | Formato | |
|---|---|---|---|
|
TESI-CHIARA-LEONORI.pdf
Open Access dal 23/10/2023
Descrizione: Tesi
Dimensione
26.25 MB
Formato
Adobe PDF
|
26.25 MB | Adobe PDF | Visualizza/Apri |
I documenti in UNITESI sono protetti da copyright e tutti i diritti sono riservati, salvo diversa indicazione.
https://hdl.handle.net/20.500.12075/1055