A review of safety test methods for new car assessment program in Southeast Asian countries

4.1 AES assessment

AES is one of the control methods for safety features presently being explored. An AES component must provide precise control on time based on suitable surroundings to address severe circumstances. AES is a significantly new technology, according to Euro NCAP, and modifications to laws are expected in 2022 to utilize its potential fully. When an expected collision is identified, AES will automatically steer to avoid accidents, which is an advantage over a safety system with AEB. Furthermore, the AES system might be integrated into an ADAS for upcoming automated or driverless cars. Several researchers have reported on the progress of their AES vehicles [23,24,48]. However, due to a lack of specialized assessments and practical methods, relatively few automated steering intervention systems are currently available. Appropriate evaluation is required to anticipate technology before it is released in the market, such as the ASEAN NCAP safety rating of a new automobile.

A change from a technology-based approach (e.g., solely testing for AES or AEB) to more scenario-based evaluations that allow for multiple sorts of interventions (e.g., braking and steering) is required [7]. Whatever technology emerges in the future, the assessment should be ready. Researchers have presented an example of a technology-based method [23,24]. An innovative ESC technique is based on a hierarchical control architecture with decision-making and motion control levels. The effectiveness of the suggested control technique for performing an emergency collision avoidance maneuver has increased. ESA devices and methods have been developed [16]. Continental, for example, a significant vehicle supplier, debuted its ContiGuard ESA system in 2010 [21]. 2 years later, a Japanese carmaker, Nissan, unveiled the concept of a self-developed helper that can drive itself in an emergency. TRW Automotive, on the other hand, has created radar and video camera-based driver assistance systems. The technology was expected to be ready for production in 2017, with applications for the 2018 model year [20].

4.2 AES demand

Many modern automobiles have safety and comfort features to address market expectations for minimizing road accidents. It is impossible to provide safety and comfort without combining steering and pedal control [49]. ADAS, lane departure warning systems, forward collision-avoidance, and adaptive cruise control are just a few of the safety innovations that have emerged as a result of these initiatives [34]. Because system failure is risky for the driver, researchers and automobile firms continue to argue the usefulness of the safety system. However, if the safety system evaluation demonstrates that it can work with a few mistakes, this is a beautiful chance to market the product and establish consumer trust [50]. Human error, such as a driver’s delayed response time owing to an unexpected presence of an impediment, causes crashes or accidents. Many collisions could be avoided or mitigated with the existing AEB technology. Although more technically challenging, AES may result in even more substantial reductions in collisions and fatalities, particularly in a single vehicle and minor overlap crashes, as well as accidents involving vulnerable road users such as bicycles and pedestrians [22].

4.3 AES control system

AES is a lateral safety mechanism that regulates steering rotation in a probable accident. Complex car models and hard math are used in the system. The main goal is to avoid colliding with the barrier by moving the steering wheel. AES control is designed to tackle collision avoidance difficulties in high-speed scenarios (highway traffic) [23,24], whereas AEB control is better suited to slow-speed settings (urban traffic) [18,21]. In high-speed conditions, the longitudinal distance to the obstruction is shorter, making AEB ineffective at preventing collisions. However, this does not imply that AES can function without AEB; instead, both systems must be used together. When a probable collision is identified, the driver must make many decisions and judgments, including “What is the obstacle type?” “Should I brake or steer at this point?” “Where can I go away?” and so on. There is so much data to analyze in a single second that the driver’s reaction time is delayed. Here ADAS, which combines the AES and AEB, assists the drivers by alerting them, braking the car, and turning the steering wheel. At first, the technology will warn the driver of impending crashes. An accident will occur if the motorist does not respond to brake until the last moment. When AEB is enabled, the vehicle speed in a lane is reduced, giving the driver more time to respond. Assuming the driver continues to fail to steer or have a reaction until the very last moment [16] and the steering torque is insufficient [49], in that circumstance, AES is engaged to prevent a collision by automatically changing lanes. The optimal braking distance that assures the AES’s effectiveness is still an unresolved problem that requires additional testing.

In Figure 3, the flow of the collision avoidance system is started by scanning the surrounding situation, such as the traffic conditions [9], road type [19], lane type [21], vehicle or pedestrian [22], static or dynamic [23], and obstacle type (big or small) [24]. The system should be concerned about the dimension in terms of obstacle type. The possibility of a rear-end collision will rise if the dimension is not evaluated. Hamid et al. [24] lengthened the obstruction using an invisible rectangle, resulting in more secure escape paths. Road, lane, and traffic data limit the best escape path alternatives. A pedestrian walkway, an incoming traffic lane, or a severely damaged road should not be considered the best escape paths [9]. The cameras, radar, and LIDAR are standard sensors used in the scanning process for determining the best escape path [16,18]. At least one sensor must produce information regarding the surroundings and the impending barrier [20]. The exchange of information between two vehicles can also improve data quality. Electronic maps may also gather environmental data over a broader region. However, they perform less precisely in tiny spaces than direct sensors on moving vehicles. The system then uses AES and AEB to follow the escape path trajectory with minimal error after choosing the best escape path.

Figure 3

Algorithm flow of collision avoidance system by applying AEB and AES to track the escape route trajectory [3].

4.4 Integration of safety assist system with AI

Vehicle technology developments such as C-ITS, the IoT, AI in cars, and autonomous driving are significant development priorities for today’s global automakers [6]. They improved vehicle safety and made vehicles more intelligent and environmentally friendly. AI was developed in three primary areas [51]: ML, neural networks (NNs), and deep learning (DL). The Scopus search platform yielded a large number of results. However, as shown in Tables 1–3, only methods relating to rear-end collisions involving automobiles were included and discussed in this article.

The fuzzy logic [34,35,36] was found on the Scopus platform for the ML approach, as shown in Table 1. Guo et al. [34] and Huang et al. [35] developed a strategy for developing lane-change decision-making utilizing fuzzy logic to assure safety. Furthermore, Tork et al. [36] proposed combining two methods: ML and NN, which is Multilayer Neural Network modified by Interval Type-2 Fuzzy Logic (MMNIT2FL). This method can handle simultaneous steering and braking operations while avoiding sudden steering angle changes. Unfortunately, this method does not consider measurement error or different road conditions, and there is no real-time evaluation of the proposed approach.

In addition, for the NN approach, Convolutional Neural Network (CNN) [37,38] and Long Short-term Memory (LSTM) [39,40] were discovered. Wang et al. [37] and Virdi [38] utilized LSTM to train the data for generating the desired steering instruction. In contrast, Bojarski et al. [39] and Hassan et al. [40] employed CNN. The data were gathered on various roadways and in real-world driving situations. However, this technique has dataset limitations and focuses more on simulation than real-world situations.

Meanwhile, the search platform identified Deep Neural Networks (DNN) [41] and Deep Reinforcement Learning (DRL) [42,43] as DL methods. DNN was proposed as a technique by Mohammed et al. [41]. On the other hand, Yoshimura et al. [42] and Moghadam and Elkaim [43] employed DRL, a combination of DL and RL. Those methods yielded inspirational learning outcomes, such as identifying unexpected events covering challenging and diverse scenarios or situations the vehicle may encounter. However, these AES approaches lack safety and comfort.

Among the previous studies, [34,37,41,42] are four papers that are most similar to the topic under discussion. The approach in these papers was vision-based, and from these four, the most similar method was reported by Yoshimura et al. [42].

Guo et al. [34] provided a unique self-driving control technique for the coordinated regulation of steer and brake mechanics in vision-based autonomous cars, as illustrated in Figure 4. This work aimed to increase the qualities of safety and riding comfort successfully. In addition, it focuses on the coupled and nonlinear characteristics of autonomous cars in emergency obstacle avoidance, as well as a nonlinear coordinated braking and steering control technique. As shown in Figure 5, the adaptive fuzzy sliding mode control method and the theory of nonlinear backstepping control are used to construct a synchronized steering and braking control system. These two methods guarantee that a closed-loop system is both global approximation stable and uniformly ultimately bounded. The modeling and experimental testing results show that the proposed control strategy improves the tracking performance of autonomous cars and improves their riding comfort and stability, even under terrible driving circumstances. Furthermore, the total suggested control system has been applied to an experimental autonomous car.

Figure 4

Typical scenarios for emergency obstacle avoidance [34].

Figure 5

Block diagram for coordination of steering and braking control [34].

In the study by Wang et al. [37], the ego-sensor vehicles may identify target-vehicle trajectories within their sensor range. The ego-vehicle forecasts if the target vehicle will switch lanes in front of it, as shown in Figure 6. There may be numerous cars in the next lanes between the ego-vehicle and its front vehicle. Therefore, these target vehicles could make independent judgments. As a result, each target vehicle’s behavior should be forecasted. The future lane change scenarios are classified one-to-one, with each case containing a target and an ego vehicle. The target vehicle is unique in each case, while an ego vehicle may exist in various circumstances. This case categorization overcomes the many-to-one dilemma between the target and ego vehicles.

Figure 6

Scenario for forecasts if the target vehicle will switch lanes in front of it [37].

The adaptive lane change prediction model for nearby cars was built by Wang et al. [37] by adding an adaptive driving choice threshold to the traditional LSTM model, as shown in Figure 7. This was done considering LSTM’s strong performance in the current lane change prediction models. The model predicts lane change behaviors using data from the ego-vehicle and the target vehicle. They demonstrate that the model achieves 93.64–97.52% accuracy for the target car in the left adjacent lane and 94.30–98.01% accuracy for the target vehicle in the right adjacent lane, which are both impressive results. Furthermore, the model they offer can operate under various driving conditions.

Figure 7

Structure algorithm of LSTM model [37].

Mohammed et al. [41] created their benchmark using the CARLA simulator. They created 19 distinct scenarios spanning the whole highway town in CARLA, as shown in Figure 8, including inverted automobiles, road collisions, and cars stopping horizontally across two highways. Their performance is also influenced by two towns: CARLA town four and CARLA town five. These cities were chosen to represent highways. They tested their model in a different town with varied scenarios and settings to discover how well it can generalize to unknown situations. For each situation, the agent begins at a set point and attempts to reach a destination without colliding with other cars, roadside barriers, or other obstacles. However, in their simulation, they have a scenario where the agent is stranded and unable to complete the scenario.

Figure 8

A few scenarios spanning the whole highway town in CARLA [41].

The feature extractor is a DNN, followed by fully connected layers to map the input. It consists of raw pixels from the camera lens and the vehicle speed in the current timestamp. The output is vehicle control of the three actions, throttle, steering angle, and brake, as illustrated in Figure 9. Their suggested approach architecture separated the four pictures by repeating the image route four times for each input image. They demonstrated that utilizing a bird’s eye view and sharing the exact parameters of the feature extraction component of the four photos in the network outperforms using a single image or not sharing the features. Furthermore, they demonstrated that the network had high generalization capabilities in many contexts.

Figure 9

Structure algorithm of CNN model [41].

Yoshimura et al. [42] suggested that a DRL controller adjusts the steering angle and avoids an emergency crash. A predefined degree of brake is applied when an emergency scenario is identified. As shown in Figure 10, there are two case scenarios (a) and (b). In scenario (a), the car swerved right to avoid the scooter and came to a complete stop. In two objects scenario (b), the identical behavior would have resulted in a pedestrian accident. Thus, they highlight that their proposed controller took an additional effort to drive left to avoid a collision with the person. It demonstrates that the suggested controller considers the probability of a secondary collision and has learned to take the appropriate actions based on the scenario to prevent collisions with both objects.

Figure 10

Emergency scenarios to adjust the steering angle and avoid a crash [42]; (a) one object scenario and (b) two objects scenario.

Yoshimura et al. validate the DRL controller [42] by modeling several traffic collision scenarios while training the controller in a simulated environment. These actions teach the DRL controller how to prepare ahead to prevent collisions in the simulated environment. The controller’s performance and stability are only assessed in a simulated environment. Validating performance in the simulation does not ensure that it will perform at the same level in the real world since there is always a difference between the scenario and the situation in the simulated environment. Therefore as a response, they additionally evaluated the DRL controller in realistic driving conditions utilizing the arrangement in a real car, as illustrated in Figure 11. The outcome demonstrated that the controller effectively generalizes to the outside world. However, this specific approach is inappropriate when a person is in the side lane since the controller’s secondary maneuver was insufficient to prevent a collision with the road user. Due to the restricted variety of training scenario simulations, this specific tendency has developed. More collision situations should be included to improve the performance further.

Figure 11

Structure algorithm of DRL in the study by Yoshimura et al. [42].

The findings show that Guo et al. [34] focus solely on riding comfort and safety, whereas Wang et al. [37], Mohammed et al. [41], and Yoshimura et al. [42] tried to focus on avoiding two or more obstacles at once. Tables 1–3 summarize additional results for other papers.

4.5 AES test methodology

Simulation and experimentation are systematic methods used to evaluate the proposed AES control. There are benefits and drawbacks to both methods. The best method for validating the driving systems is experimentation [48], where AES is used in a real-world setting. Although the outcome is accurate, the fabrication of the collision scenario is challenging. For instance, Eckert et al. [17] showed how to fabricate collision scenarios in a closed circuit. In addition to guaranteeing the driver’s safety throughout the test, the researcher had to make sure the driver could sense the real-world circumstances during crashes in the experiment. The authors control the shock caused by the unexpected presence of an obstacle by not telling the test vehicle’s driver anything about it. The experiment’s results cannot be generalized because of the small number of people that participated in the testing.

Meanwhile, multiple scenarios may be tested in a short amount of time with minimal work and risk using the simulation technique [14]. Fortunately, the trustworthiness of simulation findings highly depends on the validation and verification of models with verifiable idealization. The scenario examples should be generated by real-world data, ensuring the experimental outcomes’ reliability. As a result, as proved by Kovaceva et al. [22], using both approaches will ensure the reliability of the experimental outcome.

When doing the testing, many situations are recommended over a limited amount. Multiple situations are created by varying the surrounding circumstances (obstacle, lane, and road type). Table 4 summarizes a collection of assessment situations from the reviews of relevant literature. In Figures 12–14, several typical scenarios are represented involving (1) A vehicle approaching a stationary or moving object, (2) A sudden object appearing from the side, and (3) A sudden object from the opposing vehicle’s lane crossing the lane. The vehicle must exit the lane. However, returning to the lane is optional once a collision is avoided.

Figure 12

Typical scenarios for AES testing: static or moving obstacle in the front [2].

Figure 13

Typical scenarios for AES testing: the sudden appearance of obstacles from the adjacent lane [2].

Figure 14

Typical scenarios for AES testing: incoming obstacle crossing the lane [2].

The most critical point is that while trying to exit a lane, the framework must check that no other vehicles are present. Data reveal that loss of control is responsible for around 20% of Killed and Seriously Injured (KSI). Around 15% of all vehicle accidents involve a frontal collision with a minor overlap, while 25% of all frontal crashes involve a frontal collision [52]. This scenario accounts for around 10% of the KSI in mild overlap collisions. In contrast, KSI records 36% of vulnerable road users. As a result, considering all possible scenarios is critical and will aid AES development and implementation, particularly for manufacturers.

Avoidance occurrence and trajectory tracking error are the two types of rating methods presented. The incidence of avoidance is appropriate for evaluating the functioning of the AES. Simultaneously, trajectory tracking error is appropriate for evaluating AES control performance. In the case of the ASEAN NCAP safety rating, the scoring system based on avoidance occurrence (A_occur) is recommended since it clearly states the system’s advantage. For instance, the percentage of rear-end collisions that occur when the AES is used and the severity of injuries that occur when the AES is used to minimize rear-end collisions [9,14].

Table 5 shows the sample of a scoring sheet being discussed for inclusion into the 2021–2025 ASEAN NCAP assessment protocol under advanced safety assist technologies. Three categories have been divided on the proposed scoring sheet. The first is the fitment rating score (FRS), which indicates whether or not the AES system is available in the vehicle. The highest of one point is awarded if the vehicle’s model has AES as standard equipment. In the meantime, a vehicle model does not receive any points if it does not have AES. The A_occur is then determined through 10 trials. Higher scores indicate that the AES system is more effective at avoiding the obstacle. The proposed scoring chart also incorporates the level of comfort of the driving process. It is worth noting that the assessment criteria for the AES test are still under investigation before evaluation. With this assessment, car manufacturers are encouraged to introduce a technology that will help road users prevent road crashes [15].